<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Bitrock on Medium]]></title>
        <description><![CDATA[Stories by Bitrock on Medium]]></description>
        <link>https://medium.com/@BitrockIT?source=rss-5b07e5be1c5e------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*Fllo0y1h3KF7YQa4FPg7jA.png</url>
            <title>Stories by Bitrock on Medium</title>
            <link>https://medium.com/@BitrockIT?source=rss-5b07e5be1c5e------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 13:26:39 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@BitrockIT/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Beyond the single terminal: how to manage a fleet of Coding Agents with Orca]]></title>
            <link>https://medium.com/@BitrockIT/beyond-the-single-terminal-how-to-manage-a-fleet-of-coding-agents-with-orca-25761aa4204b?source=rss-5b07e5be1c5e------2</link>
            <guid isPermaLink="false">https://medium.com/p/25761aa4204b</guid>
            <category><![CDATA[ai-agent-development]]></category>
            <category><![CDATA[orcas]]></category>
            <dc:creator><![CDATA[Bitrock]]></dc:creator>
            <pubDate>Wed, 11 Nov 2026 10:08:19 GMT</pubDate>
            <atom:updated>2026-11-11T10:08:19.512Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/900/1*aQpvrQptucwE-ao_yFdBKQ.png" /></figure><p>Over the past few months, coding agents have moved from experimentation to everyday use in many development teams. Tools such as Claude Code, Codex, Cursor and Copilot CLI have proven reliable across a growing range of tasks: refactoring, writing tests, fixing bugs.</p><p>Once the initial phase of adoption and enthusiasm is over, however, a different problem emerges. When <strong>agents become more than one </strong>and work in parallel on the same repository, the <strong>challenge</strong> is no longer getting code out of the AI, but g<strong>overning its production</strong>: isolating tasks, keeping the main branch clean and knowing at any given moment which agent is working on which part of the codebase.</p><p>This article is the result of a series of tests carried out by several members of Bitrock’s Frontend team, who introduced Orca into their daily workflow to assess its strengths, limitations and room for extension. Below we summarize what emerged: how the tool works, how it differs from the main alternatives, and how the team adapted it to its own needs by developing an internal fork.</p><h3>Task isolation with Git Worktrees</h3><p>The key difference between a classic IDE plugin and a tool like Orca lies in the underlying architecture.</p><p>If two different agents are asked to work on the same repository in the same folder, they will get in each other’s way and both lose efficiency. Orca solves the problem at its root by leveraging Git Worktrees: every time a task is assigned or a prompt is sent, the system creates an isolated directory tree in the background.</p><p># Typical workflow with the Orca CLI</p><p>orca worktree create --branch fix/session-leak --agent claude-code</p><p>orca worktree create --branch fix/session-leak-codex --agent codex</p><p># Run the same prompt in parallel on two different models</p><p>orca fanout &quot;Find and fix the memory leak in the auth middleware&quot;</p><p>The impact on everyday work is immediate:</p><ol><li><strong>Zero local conflicts</strong>: three agents can work in parallel without any risk of overwriting the files the developer is working on.</li><li><strong>Side-by-side comparison</strong>: the same task can be sent to different models (e.g. Claude vs Codex), their diffs can be compared, and the user can evaluate the best approach and merge only the valid solution.</li><li><strong>Integration with team tasks</strong>: an Issue from GitHub or Linear can be opened and its dedicated worktree created with a single click.</li></ol><p>Claude Code also uses git worktrees, but only internally, to isolate command execution within its single terminal session. <a href="https://onorca.dev/"><strong>Orca</strong></a>, an open source Agent Development Environment (ADE) designed to coordinate multiple coding agents directly within the developer’s workflow, instead acts as <strong>a higher-level orchestrator</strong>: it automatically creates and visually manages worktrees so that multiple agents can run in parallel on different branches. This way, Orca makes it possible to compare the diffs produced by each model side by side before merging into the main branch.</p><h3>Remote monitoring: the companion mobile app</h3><p>The companion mobile app (available for iOS and Android) extends the experience beyond the terminal: it syncs in real time with local sessions or sessions on remote servers (via SSH), sending push notifications when an agent completes a task or requests confirmation (Human-in-the-loop), making it possible to monitor progress and send feedback remotely.</p><h3>Orca and the alternatives on the market</h3><p>Orca enters a market that is booming right now, but with a very specific approach. If we are evaluating how to structure our Developer Experience, the main options fall into three categories:</p><h4>1. T3 Code vs Orca</h4><p>T3 Code is another very solid visual orchestrator. The main difference lies in its approach: T3 Code focuses on centralizing chat sessions and multi-provider threads in a single interface, while Orca (and therefore Forca) puts much more emphasis on system-level isolation through Git Worktrees, mobile control and support for agents running on remote machines via SSH.</p><h4>2. OpenHands</h4><p>OpenHands takes an end-to-end approach: it provides the agent, planning, the sandbox and the execution environment. Orca follows the opposite logic and positions itself as an external coordination layer. It does not replace the CLIs or subscriptions already in use (Claude Code, Codex and others), but adds a layer of management and orchestration on top of them.</p><h4>3. Workflow orchestrators (Temporal, LangGraph)</h4><p>These address a different need: tools like Temporal or LangGraph are used to write and run AI workflows inside applications in production. Orca, on the other hand, helps us developers write code faster during the development phase.</p><p>To sum up:</p><p><strong>Orca</strong></p><ul><li>Strength: Fleet management via Git Worktrees, simple and intuitive UI, mobile app</li><li>Usage model: BYOA (Bring your own CLI agents)</li><li>Weakness: Learning curve for advanced Git</li></ul><p><strong>T3 Code</strong></p><ul><li>Strength: Single interface for threads and prompts</li><li>Usage model: Freemium / BYO-Key</li><li>Weakness: Less focused on remote Worktrees</li></ul><p><strong>OpenHands</strong></p><ul><li>Strength: End-to-end autonomous system</li><li>Usage model: Open source with sandbox</li><li>Weakness: Heavier to run locally</li></ul><h3>Forca: a fork of Orca</h3><p>Since the project is released under the MIT license, Orca’s real advantage is that you are not bound by the original team’s design choices. If something your workflow needs is missing, you build it.</p><p>That is exactly what <a href="https://www.linkedin.com/in/davide-ghiotto/">Davide Ghiotto</a>, Senior Front-end Engineer, did with the help of <a href="https://www.linkedin.com/in/elhosain-halili-155119221/">Elhosain Halili</a>, Front-end Engineer. Using Orca every day for automated programming, they missed a lean, visual way to manage branches and stashes in the style of VS Code.</p><p>Instead of adapting or opening a feature request and waiting, you can work on the source code with the help of AI and create your own custom fork: “<a href="https://www.linkedin.com/posts/davide-ghiotto_altro-modo-di-usare-lai-personalizzatevi-ugcPost-7518252438294831104-39YY"><strong>Forca</strong></a>“. In a short time, they added:</p><ul><li><strong>A native branch switcher in Source Control</strong>, complete with quick commands and sync indicators (pull/push counters).</li><li><strong>Advanced Stash management,</strong> with dedicated submenus for apply, pop and drop, plus an automatic “Stash &amp; Switch” function when a branch switch is blocked.</li><li><strong>Signed macOS builds</strong>, so that anyone on the team can install and use the modified version right away.</li></ul><p>This is a great example of how AI is changing the way we approach open source: we no longer simply use tools as they are provided, but shape them around the team’s real needs with minimal effort.</p><h3>Conclusions</h3><p>Orca does not make agents smarter: if a model produces wrong solutions or hallucinations, it will keep doing so. Its value lies elsewhere, namely in reducing the cognitive load of context-switching and in making it possible to treat coding agents for what they are today: a group of assistants to whom well-defined tasks can be delegated, to be checked and reviewed with the right tools before changes reach the main codebase.</p><p>Our Front-end team’s experience suggests that maturity in adopting coding agents is not measured by the number of agents in use, but by the quality of the process that governs them. Task isolation, structured comparison between alternative solutions and a human checkpoint before merging are practices that go beyond any single tool, and that every team should consider when integrating AI into its development cycle.</p><h3>Useful links</h3><ul><li>Orca: <a href="https://www.onorca.dev/">https://www.onorca.dev</a></li><li>LinkedIn post about Forca: <a href="https://www.linkedin.com/posts/davide-ghiotto_altro-modo-di-usare-lai-personalizzatevi-ugcPost-7518252438294831104-39YY">https://www.linkedin.com/posts/davide-ghiotto_altro-modo-di-usare-lai-personalizzatevi-ugcPost-7518252438294831104-39YY</a></li><li>Forca GitHub repo: <a href="https://github.com/davide97g/orca">https://github.com/davide97g/orca</a></li></ul><p>Main Author: <a href="https://www.linkedin.com/in/danielzotti/">Daniel Zotti</a>, Team Leader &amp; Tech Leader Frontend @ Bitrock</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=25761aa4204b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Self-Hosted PaaS: What Digital Sovereignty Costs in Practice]]></title>
            <link>https://medium.com/@BitrockIT/self-hosted-paas-what-digital-sovereignty-costs-in-practice-50e697cc84da?source=rss-5b07e5be1c5e------2</link>
            <guid isPermaLink="false">https://medium.com/p/50e697cc84da</guid>
            <category><![CDATA[digital-sovereignty]]></category>
            <category><![CDATA[coolify]]></category>
            <category><![CDATA[frontend]]></category>
            <category><![CDATA[paas]]></category>
            <dc:creator><![CDATA[Bitrock]]></dc:creator>
            <pubDate>Tue, 22 Sep 2026 15:26:01 GMT</pubDate>
            <atom:updated>2026-09-22T15:26:01.336Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/900/1*EobIQvDXII7_prC8dJd0HA.png" /></figure><p>Sovereignty is easy to declare and expensive to operate. Last spring we built <strong>Fortitudex</strong>, the collectible card app for the <strong>Fortitude Group Convention</strong>: three deployable units, a few hundred concurrent users, and a fixed date on the calendar.</p><p>The quickest route was three American SaaS platforms and a weekend of setup. We took a slower one instead, and ran the whole project on <strong>a self-hosted PaaS</strong>, on infrastructure we control.</p><h3>The gap between a policy and a git push</h3><p><strong>The Fortitude Group has been explicit about digital sovereignty as a strategic position</strong>: autonomy and control over where workloads run, without giving up innovation speed. That conversation usually happens at the level of critical systems, the ones that get audited, mapped and put in front of a compliance committee.</p><p>Dependency tends to accumulate somewhere else. It builds up in the internal project that nobody audits: the demo, the small internal tool, the event app. Nobody writes a data-residency assessment for those. Somebody opens a laptop, connects a repository to a hosting platform, and by the end of the afternoon the code, the build logs, the database and the outbound email of a company project sit on three vendors in two jurisdictions, under a click-through agreement.</p><p>Do that for every side project in an engineering organisation and the result is your actual posture, which no policy document describes.</p><h3>What we were not willing to give up</h3><p>We did not set out to “go on-premise”. We set out to keep the developer experience that made those SaaS platforms the default in the first place, and to change only where it runs.</p><p>The list is short, and we treated it as non-negotiable:</p><ul><li>A preview environment for every pull request, on its own URL</li><li>Deploy on push, with no manual step</li><li>Certificates and DNS handled by the platform rather than by a person</li><li>Rollback that takes less time than explaining the incident</li></ul><p>This matters more than it sounds. A sovereignty strategy that slows the team down does not survive first contact with a deadline. It gets quietly bypassed, and the organisation ends up paying the infrastructure bill and keeping the shadow IT anyway. For the strategy to hold, the sovereign option has to be the easier one at the moment a developer chooses.</p><p>Preview environments are the clearest case. On a metered platform every pull request carries a price, teams start to feel it, and they open fewer of them. On a machine you are already paying for, one more preview costs nothing, so the practice gets used the way it was meant to be. The billing model stops making engineering decisions for you.</p><h3>Why a platform, and not just a server</h3><p>Owning the hardware is the easy half. We could have taken the same machine, installed Docker, written a reverse proxy configuration by hand, scripted the certificate renewals and wired up a deploy hook. Every step of that is well documented and none of it is hard.</p><p>It is also how a team ends up with a server that only one person can safely touch. The configuration lives in that person’s head and in files on that host. Onboarding a colleague means booking a walkthrough. Preview environments never get built, because each one is manual work that somebody has to want to do.</p><p>A self-hosted PaaS gives you the shape of a managed product on a machine you control: a web interface, containers, automatic TLS, git-triggered deploys, a preview environment per pull request. The property we were really buying is that a second person can operate it without a handover meeting. On a project with a fixed date and more than one contributor, that decided it.</p><h3>Coolify or Dokploy?</h3><p>Two projects dominate this space, and they are solving slightly different problems.</p><p><strong>Coolify</strong></p><p><strong>Dokploy</strong></p><p>First release</p><p>2022, mature with a large community</p><p>April 2024, young and fast-moving</p><p>Licensing</p><p>Open source, self-hosted free with no feature limits</p><p>Open source, self-hosted free</p><p>Scope</p><p>Broad: 280+ one-click services, multi-server, native MCP server since v4.0 (May 2026)</p><p>Deliberately narrow: Docker-first, minimal surface</p><p>Interface</p><p>Feature-dense, more to learn</p><p>Cleaner and faster to read</p><p>Preview environments</p><p>Per-pull-request previews, first-class</p><p>Present, less developed</p><p>Recent trajectory</p><p>Structural releases</p><p>Incremental</p><p>Best fit</p><p>A substitute for a managed platform</p><p>The thinnest usable layer over Docker</p><p>We chose Coolify for two reasons, neither of which was a feature count. Per-PR preview environments were on our non-negotiable list, and Coolify’s implementation of them was the more complete one. The second reason was maturity, which in self-hosted software is a proxy for something specific: when the platform breaks, your support channel is other people who have hit the same thing. A four-year-old project with a large community can answer questions a newer one has not been asked yet.</p><p>That is a judgement about fit for our case. A team that wants a small, legible layer over Docker Compose and nothing more would be right to pick the narrower tool, and the smaller surface area is a real advantage when you are the person on call for it.</p><h3>What it cost us in surprises</h3><p>What you take on here is a set of standing obligations, and they need an owner with a name. Backups, patching, certificate renewals and capacity headroom become recurring items on somebody’s calendar. If nobody owns them, they have not gone away; they have moved onto a risk register that nobody reads.</p><p>None of the following appears in the documentation, and all three would change how we set this up a second time.</p><p><strong>Configuration drifts out of the repository.</strong> The platform’s own state, meaning services, environment variables and routing, lives in its web interface and not in your git history. For a team used to keeping infrastructure next to code, that is a step backwards: the application is versioned and the platform is not. We can live with it at this size. At ten times the size we could not.</p><p><strong>Deploys are less isolated than they look.</strong> Production, the development environment and every open pull request share one host and one entry point, which makes them neighbours. A deploy in one of them can be briefly visible in another. On a demo that is a shrug. It is also the reason we would not put anything with an availability commitment on this setup as it stands today.</p><p><strong>Every managed service you drop becomes a standing obligation.</strong> We started on a managed database and later moved to self-hosted Postgres on the same platform. The migration was ordinary work and took an afternoon. What changed permanently was the accounting: backup verification and upgrade windows belong to us now, and they will still belong to us in three years.</p><h3>Conclusions</h3><p>The finding worth reporting from Fortitudex is not really about Coolify, which is good software but replaceable. It is that the trade-off most organisations assume (<strong>sovereignty on one side and developer experience on the other</strong>) did not apply at our scale. We kept preview environments per pull request, deploy on push and automatic TLS, and we moved the whole thing onto hardware we control, inside a project that lasted a few weeks.</p><p>None of this is a recommendation. It is one project, with a small team, an internal audience and a deadline we controlled, and those properties did most of the work. The operational load is the part we did not escape. Digital sovereignty is not a purchase that completes. It is a capability that has to be staffed and kept staffed, and the machine is the least demanding part of it. Where we have seen this go wrong, the machine was planned for and the time to look after it was not.</p><p>Main Author: <a href="https://www.linkedin.com/in/andreamazza1/">Andrea Mazza</a>, Senior Front-end Developer @ Bitrock</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=50e697cc84da" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Shadow AI and AI Sprawl: How to Centralize LLM Governance at the Infrastructure Level]]></title>
            <link>https://medium.com/@BitrockIT/shadow-ai-and-ai-sprawl-how-to-centralize-llm-governance-at-the-infrastructure-level-9795c91439a6?source=rss-5b07e5be1c5e------2</link>
            <guid isPermaLink="false">https://medium.com/p/9795c91439a6</guid>
            <category><![CDATA[ai-sprawl]]></category>
            <category><![CDATA[shadow-ai]]></category>
            <dc:creator><![CDATA[Bitrock]]></dc:creator>
            <pubDate>Wed, 02 Sep 2026 09:38:51 GMT</pubDate>
            <atom:updated>2026-09-02T09:38:51.067Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/900/1*4WkCHUOJmtQgPON7s9p8RA.png" /></figure><p><strong>AI Sprawl</strong> refers to the uncontrolled proliferation of disconnected AI tools and models within a company. <strong>Shadow AI</strong>, on the other hand, is the unauthorized use of Generative AI applications by employees, outside IT control. Both phenomena expose enterprise companies to serious security vulnerabilities, hidden cloud costs and legal non-compliance risks.</p><h3>The Hidden Risks of AI in Enterprise Environments</h3><p>Generative AI tools have become essential infrastructure for everyday productivity. This efficiency gain is real and represents an undeniable competitive advantage. However, <strong>the ease of access to these technologies has removed every barrier to entry.</strong></p><p>Today, the “friction to value” of an LLM is close to zero; all an employee needs is a personal browser and a login account to bypass every corporate security firewall. This immediate accessibility has given rise to two distinct but interconnected phenomena that are undermining the foundations of IT Governance: <strong>Shadow AI</strong> and <strong>AI Sprawl</strong>.</p><p>To limit these risks, companies face a true structural dilemma. On the one hand, blocking AI adoption means condemning the company to operational obsolescence. On the other, letting teams act in complete autonomy is creating a <strong>fragmented and inherently insecure ecosystem</strong>.</p><p>The key question today is therefore whether you have a <strong>scalable governance framework ready to support the solutions already in use</strong>. How can the speed of experimentation demanded by the business be balanced with the control needed to protect company assets?</p><h3>Understanding Shadow AI and AI Sprawl</h3><p>First of all, it is essential to accurately map the differences between these two trends: although <strong>both stem from a central governance deficit</strong>, their vectors and impacts differ significantly.</p><h3>Shadow AI: Unauthorized Use at the User Level</h3><p>Shadow AI occurs when individual employees or entire business departments <strong>adopt and feed external AI tools without formal approval from the IT department or without following precise guidelines</strong>.</p><p>The driver of this behavior is almost never malicious; on the contrary, it stems from the simple attempt to work faster or more easily.</p><p>When a user copies and pastes company information into a public chatbot, that information leaves the company’s controlled environment to be processed by external services with unknown storage policies. IT departments thus lose all visibility into the content of the HTTPS requests sent to providers such as OpenAI, Anthropic or Google.</p><h3>AI Sprawl: Infrastructure Proliferation</h3><p>AI Sprawl, by contrast, is the <strong>uncontrolled proliferation of AI tools, APIs and models integrated within company systems</strong>. This phenomenon occurs when development teams adopt solutions in a fragmented way, <strong>without a central architecture strategy.</strong></p><p>Thanks to modern APIs, what once required months of work by a team of data scientists can now be prototyped by a single developer with a few API calls and a pre-trained model. While <strong>this agility accelerates release cycles, it also masks a macroscopic problem: the duplication of capabilities</strong>.</p><p>Without common guidelines, different business units end up building nearly identical micro-services using different providers, separate credentials and independent billing cycles. Every quick prototype brings with it long-term maintenance, updates and monitoring. <strong>In an ungoverned context, engineering overhead stops creating value and turns into the management of a chaotic library of overlapping tools.</strong></p><h3>Security, Compliance and Hidden Costs</h3><p>The risks arising from the combination of Shadow AI and AI Sprawl go far beyond operational inefficiency, moving directly into the territory of <strong>legal liability</strong> and <strong>data integrity</strong>.</p><h3>Sensitive Data Exposure</h3><p>The main and most immediate risk is so-called <strong>data leakage</strong>. When employees interact with public tools, the data entered into prompts can be stored, logged and even used by vendors for the future training of their models.</p><p>Real-world scenarios include developers pasting portions of proprietary code containing hardcoded credentials or internal IP addresses to fix a bug; legal teams using online LLM assistants to summarize confidential contracts related to mergers and acquisitions; data analysts uploading entire customer databases with plain-text email addresses to extract charts and commercial insights.</p><p><strong>The scattering of API keys</strong> across scripts, notes or instant messaging tools <strong>dramatically widens the corporate attack surface, making the infrastructure hard to defend.</strong></p><h3>The Compliance Paradox and Data Persistence</h3><p>A crucial aspect concerns the persistence of derived data. When a team feeds a model for fine-tuning or to implement RAG logic, <strong>proprietary data does not vanish when the original source is deleted</strong>. Embeddings, insights and derived weights persist stably within the AI ecosystem.</p><p>This creates an <strong>irreconcilable conflict with current regulations</strong>, particularly the GDPR. If a customer exercises their right to be forgotten, simply deleting the records from the company CRM is no longer enough. If that data has been ingested by a shadow or untracked AI tool, it lives on in the form of entries in a vector database or patterns within a fine-tuned model.</p><p><strong>Demonstrating compliance under these conditions becomes an operation of high technical complexity and extremely high sanction risk.</strong></p><h3>Cloud Cost Explosion and Fragmentation</h3><p>In the absence of centralized monitoring, <strong>quantifying overall Generative AI spending becomes impossible</strong>. The company ends up paying fragmented “retail” rates across dozens of different accounts, losing the opportunity to negotiate enterprise commercial agreements based on the consolidation of total token volumes.</p><h3>Overcoming Silos: An Analysis of the Approaches</h3><p>The immediate reaction of many companies to these risks is to impose rigid bans. However, the history of cloud computing has taught us that absolute blocks simply push usage underground, exacerbating the problem. <strong>The real solution lies in evolving the infrastructure.</strong></p><p>The table below provides a structural analysis of the differences between fragmented management (silos) and centrally governed, strategic management:</p><p><strong>Technical CharacteristicSilos / Shadow AIUnified Centralized Approach (AI Gateway)Cost Visibility</strong>Fragmented, unknown and spread across separate retail billing.Real-time token monitoring and consolidation of budgets per team.<strong>Data Privacy &amp; GDPR</strong>High risk of personal data leaks; no PII scrubbing.Centralized PII filters and automated masking at the infrastructure level.<strong>Team Agility</strong>High in the short term, but chaotic and burdened by future technical debt.High and scalable, protected by clear architectural guardrails.<strong>Compliance &amp; Audit</strong>Reactive, complex and performed through “technical archaeology”.Proactive, simplified and supported by centralized log records.</p><h3>Bitrock’s Approach</h3><p>At Bitrock we believe that AI Governance should not be a brake on operations, but the fundamental enabling factor for scaling business solutions securely and efficiently, as we demonstrated in our recent article.</p><p>The architectural model we propose also relies on the <a href="https://radicalbit.ai/ai-gateway/"><strong>Radicalbit AI Gateway</strong></a>, a product in the <a href="https://fortitudegroup.it/en">Fortitude Group</a> portfolio, which acts as a reverse proxy intercepting and orchestrating every interaction between business applications and LLM models.</p><p>In detail, our approach structures the intervention in incremental phases that do not interrupt teams’ current workflows:</p><ol><li><strong>Auditing and verification:</strong> identify which API keys and SaaS contracts are active across the various departments to map the existing landscape.</li><li><strong>Use case mapping:</strong> classify the required functions (text summarization, data analysis) to surface and eliminate real overlaps and ensure the risk classification required by the EU AI Act.</li><li><strong>Technical centralization:</strong> route all LLM traffic through a centralized AI Gateway as the single secure access point.</li><li><strong>Policy enforcement:</strong> apply security filters, PII masking and compliance rules automatically at the infrastructure level, without touching application code.</li></ol><p>This solution fully abstracts the model provider, delivering immediate benefits: <strong>it allows you to switch LLMs without touching a single line of application code, guarantees real-time observability of costs and consumed tokens, and automatically applies security filters and sensitive-data anonymization policies before that data leaves the company perimeter.</strong></p><p>Furthermore, thanks to advanced capabilities such as intelligent caching and dynamic routing, the AI Gateway reduces latency and operating costs, providing Model Fallback and Load Balancing mechanisms to eliminate business downtime.</p><h3>Conclusion</h3><p>The spread of Shadow AI and AI Sprawl is a natural consequence of a powerful technology growing faster than traditional governance.</p><p><strong>Centralizing access to models through an AI Gateway makes it possible to regain visibility over data flows and optimize cloud costs without undermining developer agility.</strong></p><p>Don’t let the uncontrolled proliferation of AI put your company’s security and compliance at risk. <a href="https://bitrock.it/contacts"><strong>Contact</strong></a><strong> us to find out how we can help you govern your AI stack efficiently and securely.</strong></p><h3>FAQ</h3><p><strong>How do Shadow AI and AI Sprawl undermine data protection strategies?</strong></p><p>The core risk lies in the invisible leakage of data. When an employee uses unauthorized tools, giving rise to the phenomenon known as Shadow AI, the prompts they send often contain proprietary code or sensitive data that external vendors may reuse to train their own algorithms. From the AI Sprawl perspective, the fragmented integration of unmonitored APIs makes it impossible to trace where derived data, embeddings or fine-tuned model weights reside, violating compliance requirements such as the GDPR or the EU AI Act.</p><p><strong>Why do traditional security solutions fail against Shadow AI?</strong></p><p>Traditional cybersecurity thinks in terms of logical perimeters. With Generative AI, this model collapses because friction has been eliminated: if the company blocks a portal on the corporate network, users instantly move to personal devices or use alternative APIs to automate their work. Bans do not stop adoption; they only make it invisible to IT. The new paradigm must instead intercept and filter the communication flows between applications and models through a single access point.</p><p><strong>How can a company map and quantify the risk of ungoverned AI without blocking team agility?</strong></p><p>The answer lies in adopting a transparent infrastructure framework. Instead of imposing long bureaucratic approval processes that slow down innovation, the company introduces a centralized access point, such as Radicalbit’s AI Gateway, which unifies traffic, automatically masks sensitive data and standardizes credentials. This way, developers retain full agility in testing models, while the organization gains the structural visibility needed for compliance.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9795c91439a6" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI Won’t Replace Developers: Why Skills and Soft Skills Still Matter]]></title>
            <link>https://medium.com/@BitrockIT/ai-wont-replace-developers-why-skills-and-soft-skills-still-matter-e8bf89c84866?source=rss-5b07e5be1c5e------2</link>
            <guid isPermaLink="false">https://medium.com/p/e8bf89c84866</guid>
            <category><![CDATA[mobile-app-development]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ai-mobile-app-development]]></category>
            <dc:creator><![CDATA[Bitrock]]></dc:creator>
            <pubDate>Tue, 01 Sep 2026 10:31:25 GMT</pubDate>
            <atom:updated>2026-09-01T10:31:25.714Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/900/1*CR1_d8MGkndvFhD8YEMWtg.png" /></figure><p><strong>Third in a three-part series inspired by the </strong><a href="https://www.linkedin.com/event/manage/7460258465408573440/">roundtable</a><strong> Bitrock organized with nine Mobile Development and AI Professionals.</strong></p><p>Having looked at how AI has made its way into mobile development and how it behaves across different frameworks, one question still looms largest: <strong>what does this actually change for people and for businesses</strong>?</p><p>On this topic, the Bitrock roundtable offered its deepest reflections — along with a few honest concerns.</p><h3>Risk Number One: The “Yes-Man” Trap</h3><p>The biggest danger isn’t technological — it’s an attitude. Get used to clicking “yes” on every AI suggestion without understanding what’s happening behind the scenes, and you turn into what the group called a “yes-man.” That behavior, warns moderator Samantha Giro, Mobile Manager at Bitrock, can become the real source of bugs: not seeing what you’re doing, or why, is a serious problem.</p><p>Carlo Lucera, Flutter Team Lead at Pivotal Technologies, ties this to understanding “why” something works: if a junior gets used to only reviewing generated code, then when the AI hallucinates or suggests an outdated pattern, “they don’t have the critical tools to catch it.” Human expertise on the framework, he concludes, remains the difference “between an app that holds up and one that breaks at the first system update.”</p><h3>How Do You Train Juniors in Such a Fast-Moving World?</h3><p>This is where the deepest concern surfaces. Alberto Dallaporta, founder of the consulting firm Novalab, raises an issue every company has to reckon with: “It’s genuinely hard to figure out how to grow people, when in six months they do a tenth of what we used to do in six months.” His advice to newcomers today is blunt: start from the basics of native code and platform knowledge anyway — permissions, platform-specific behaviors, the differences between iOS and Android — while cultivating curiosity and soft skills.</p><p>Emanuele Maso, an Android developer at Bitrock, echoes this, urging people to never let go of the drive to understand: “You still need that curiosity. If there’s a piece of code I don’t get, I ask the AI to explain it to me, because I want to understand things regardless.” For those just starting out, he suggests going deep on one technology, getting really good at it, and only then branching out: “The rest comes almost for free after that.”</p><p>Federico Nessi, an iOS developer at Bitrock, adds a valuable nuance. Before AI, learning meant spending hours tracking down why a bug happened — and that effort is what built expertise. “AI can shortcut that search, but my advice is to keep your curiosity alive: don’t just accept the solution, stay curious about why something works the way it does. That’s what gives you a real picture of how the machine works underneath.”</p><h3>Client Expectations: Watch Out for the “Everything Free and Instant” Mindset</h3><p>There’s also a chapter that touches the business side directly. Alberto Dallaporta, who deals with clients every day, describes a common and risky expectation: “A lot of people read sensationalist headlines and assume an app today costs half as much and takes half the time.” The reality is more nuanced: AI speeds up writing code, but that’s only a fraction of the project. “Requirements analysis, architecture, testing, UX design, and maintenance take the exact same effort as before, if not more, precisely because you need to validate what the AI produces.”</p><p>It’s the principle insiders sum up as “garbage in, garbage out.” Mauro Marinello, Data Scientist and Product Owner at <a href="https://radicalbit.ai/">Radicalbit</a> (a solution within the <a href="https://fortitudegroup.it/en">Fortitude Group </a>product portfolio) puts it well: AI doesn’t “think” about the business goal, it generates tokens on a probabilistic basis. “The role of the Product Owner, the analyst, and the architect becomes even more central: we’re the ones who have to set the boundaries the agent operates within. If the requirements are ambiguous, the AI will generate ambiguous code.”</p><p>One real example says more than a thousand explanations: the roundtable brought up the case of an influencer who had an AI “build” their own social network, only to end up with a hefty fine for mishandling privacy and GDPR compliance. Writing software — mobile or web — also means managing the lifecycle, data security, resource optimization, and regulatory compliance: skills AI doesn’t replace.</p><h3>The Developer of the Future: A Director, Not a Keyboard Operator</h3><p>If one line sums up the whole discussion, it’s from Marco Riva, AI engineer at Fortitude Group: “The focus needs to shift from pure code syntax to logic and architecture. Memorizing a language’s syntax matters less and less, because that’s exactly what AI does better and faster than we do. The developer of the future will increasingly be a director and a validator, and less and less a keyboard operator.”</p><p>A clear-eyed optimism prevails when it comes to the near future too. Stefano Mondino, an iOS developer at Synesthesia, is betting AI integration will become so natural it disappears entirely: “We won’t even talk about ‘using AI’ anymore — it’ll just be the standard way of coding, the same way today we don’t say ‘I’m using autocomplete.’”</p><h3>The Takeaway</h3><p>Mobile development is still here and it’s here to stay. AI won’t replace developers, because a well-built product is made by skilled people. The tool is incredibly powerful — but, as someone put it during the discussion, if you don’t know what to type, it just helps you break things faster.</p><p>The future rewards those who know how to ask the right questions, understand the why, and keep their curiosity alive — qualities no model, for now, can generate on our behalf.</p><p><strong>Three-part series — </strong><a href="https://www.youtube.com/watch?v=PURSkB7hjt8"><strong>Watch the Roundtable again</strong></a></p><p>Speakers: <a href="https://www.linkedin.com/in/federico-monti-a4b29715b/">Federico Monti</a> (MOLO17), <a href="https://www.linkedin.com/in/stefanomondino/?locale=en">Stefano Mondino</a> (Synesthesia), <a href="https://www.linkedin.com/in/clucera/">Carlo Lucera</a> (Pivotal Technologies), <a href="https://www.linkedin.com/in/marco-gomiero/">Marco Gomiero</a> (Airalo / Google Developer Expert), <a href="https://www.linkedin.com/in/fnessi/">Federico Nessi </a>(Bitrock), Emanuele Maso (Bitrock), <a href="https://www.linkedin.com/in/adallaporta/">Alberto Dallaporta </a>(Novalab), <a href="https://www.linkedin.com/in/mauromariniello/">Mauro Marinello</a> (Fortitude Group), <a href="https://www.linkedin.com/in/marco-riva-098a72205/">Marco Riva </a>(Fortitude Group).</p><p>Moderator: <a href="https://www.linkedin.com/in/samantha-giro-28106ba7/">Samantha Giro</a>, Mobile Manager at Bitrock.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e8bf89c84866" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why Does AI “Prefer” React Native? Frameworks and Tech Choices in the Age of Agents]]></title>
            <link>https://medium.com/@BitrockIT/why-does-ai-prefer-react-native-frameworks-and-tech-choices-in-the-age-of-agents-b64206bc6b41?source=rss-5b07e5be1c5e------2</link>
            <guid isPermaLink="false">https://medium.com/p/b64206bc6b41</guid>
            <category><![CDATA[mobile-app-development]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Bitrock]]></dc:creator>
            <pubDate>Tue, 28 Jul 2026 09:17:14 GMT</pubDate>
            <atom:updated>2026-07-28T09:17:14.515Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/900/1*e94ZO2jLmLvHoH9E3lZgMg.png" /></figure><p>Second in a three-part series inspired by the <a href="https://www.linkedin.com/event/manage/7460258465408573440/">roundtable</a> Bitrock organized with nine Mobile Development and AI Professionals.</p><p>It only takes a small experiment to spot a curious pattern. <strong>Samantha Giro, Mobile Manager at Bitrock</strong>, asked an AI to generate an app without specifying which architecture to use, then repeated the test across different platforms. The result left her puzzled: for cross-platform work, the agent almost always suggested React Native. Not Flutter, not Kotlin Multiplatform. Why?</p><p>That question kicked off one of the most interesting threads at the Bitrock roundtable on the relationship between AI and mobile development. In this second blog post, we get into the details of tech choices: <strong>how AI behaves with different frameworks, and what that means in practice for people designing apps</strong>.</p><h3>It’s a Numbers Game (and a Training Data Game)</h3><p>The most widely shared explanation is as simple as it is telling: AI knows best whatever has the most code written about it. Stefano Mondino, an iOS developer at Synesthesia, nails it: “It comes down to numbers. React Native draws on a JavaScript developer pool that’s massively bigger than, say, Flutter’s, which starts from a less widely known language like Dart.”</p><p>Marco Riva, AI engineer at Fortitude Group, confirms this from the model-building side: without precise instructions or a preliminary web search, the model heads straight for whatever its training data covers best — which means JavaScript and React Native.</p><p>Carlo Lucera, Flutter Team Lead at Pivotal Technologies, adds an original take rooted in where these tools came from: many were built for the web, where getting an instant app preview meant using technologies that run in the browser. “That’s where the preference came from, and it stuck around, becoming a cornerstone of agentic coding.”</p><h3>Every Framework Has Its Own Pitfalls</h3><p>AI’s behavior shifts quite a bit from one technology to the next, and knowing where those limits lie is simply part of the job.</p><p>On <strong>Flutter</strong>, Carlo Lucera points less to glaring errors than to slightly dated habits: “Every so often it uses outdated conventions, like declaring the type of a constant when it’s no longer needed. They’re not bugs, just extra clutter to clean up. Reliability-wise, though, we’re at close to 90% correct.”</p><p>On <strong>Kotlin Multiplatform</strong>, Marco Gomiero acknowledges the training data is thinner, “but with the right pointers and a few skills or guardrails, it actually works out pretty well — the key is giving it precise instructions.”</p><p>On <strong>iOS</strong>, Federico Nessi noticed a stark difference between projects: on a messy implementation, the AI tended to lose its way in the logic flow, whereas on a clean, widely-adopted architecture like TCA it was “extremely precise.” His take: patterns implemented consistently across the whole community generate sharp, well-defined skills — and that lines up perfectly with how an AI reasons.</p><p>On <strong>Android</strong>, Federico Monti describes refactors complicated by outdated library references: “Getting to the final version took a lot of prompting and a lot of digging online.” And he flags a real risk with Jetpack Compose: AI also draws on mountains of dated code built on XML or older Compose versions, which can saddle you with technical debt “from day one” if you don’t have the expertise to catch it.</p><h3>The Real Difference-Maker: Project Structure</h3><p>On one point, everyone agrees: the more organized the project, the better the outcome. Stefano Mondino sums it up with a practical, often-underrated detail — folder structure. “With a really solid setup and well-organized folders, the folder name itself becomes context for the AI. Claude works like a dream.” Federico Monti makes the same point from another angle: sensible file and folder names, internal documentation, and comments cut down maintenance work dramatically down the line.</p><h3>Native or Hybrid? The Question Itself Is Changing</h3><p>The inevitable provocative question: if agents can generate solid native code, does hybrid development still make sense? The answers converge less on a clear winner than on a shift in perspective.</p><p>Marco Gomiero suggests staying flexible: “The real shift is not getting fixated on native versus multiplatform, but using whatever the agent handles best. It’s a mindset change too.” Stefano Mondino brings the focus back to people: “If I’ve got a highly skilled iOS team and a highly skilled Android team, I’d rather help them collaborate better and stay native, instead of taking on the risk of a drastic architecture change.” In the end, the choice comes down to time, budget, the quality bar, and the team you’ve got.</p><p>And looking ahead? Mondino is betting on a fairly stable division of labor: native will hold onto its share for premium, high-performance apps with deep hardware integration, while cross-platform frameworks like React Native and Flutter — increasingly powered by AI agents — will dominate everyday utility apps.</p><h3>The Bottom Line</h3><p><strong>AI isn’t neutral when it comes to frameworks</strong>: it reflects what it’s learned, and what it’s learned depends on how much code is out there to learn from. Knowing this lets you make informed choices, give the right instructions, and structure projects so the agent can do its best work.</p><p>In the third and final piece, we’ll tackle the trickiest question of all: what remains irreplaceably human, from growing junior developers to managing client expectations.</p><p><strong>Three-part series — </strong><a href="https://www.youtube.com/watch?v=PURSkB7hjt8"><strong>Watch the Roundtable again</strong></a></p><p>Speakers: <a href="https://www.linkedin.com/in/federico-monti-a4b29715b/">Federico Monti</a> (MOLO17), <a href="https://www.linkedin.com/in/stefanomondino/?locale=en">Stefano Mondino</a> (Synesthesia), <a href="https://www.linkedin.com/in/clucera/">Carlo Lucera</a> (Pivotal Technologies), <a href="https://www.linkedin.com/in/marco-gomiero/">Marco Gomiero</a> (Airalo / Google Developer Expert), <a href="https://www.linkedin.com/in/fnessi/">Federico Nessi </a>(Bitrock), Emanuele Maso (Bitrock), <a href="https://www.linkedin.com/in/adallaporta/">Alberto Dallaporta </a>(Novalab), <a href="https://www.linkedin.com/in/mauromariniello/">Mauro Marinello</a> (Fortitude Group), <a href="https://www.linkedin.com/in/marco-riva-098a72205/">Marco Riva </a>(Fortitude Group).</p><p>Moderator: <a href="https://www.linkedin.com/in/samantha-giro-28106ba7/">Samantha Giro</a>, Mobile Manager at Bitrock.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b64206bc6b41" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI in Mobile Development: From Skeptics to Believers]]></title>
            <link>https://medium.com/@BitrockIT/ai-in-mobile-development-from-skeptics-to-believers-b37e2155bd41?source=rss-5b07e5be1c5e------2</link>
            <guid isPermaLink="false">https://medium.com/p/b37e2155bd41</guid>
            <category><![CDATA[mobile-app-development]]></category>
            <category><![CDATA[mobile]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Bitrock]]></dc:creator>
            <pubDate>Wed, 15 Jul 2026 15:03:32 GMT</pubDate>
            <atom:updated>2026-07-15T15:03:32.967Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/900/1*1B3FBzWIZGPKCRYNUD6TAg.png" /></figure><p>First in a three-part series inspired by the<a href="https://www.linkedin.com/event/manage/7460258465408573440/"> roundtable</a> we organized with nine Mobile Development and AI Professionals.</p><p>For many mobile developers, there’s a specific moment when <strong>AI stopped being a curiosity and turned into part of the job.</strong> Stefano Mondino, an iOS developer at Synesthesia for the past fourteen years, tells the story with a hint of irony: “Back in December, I never imagined I’d end up at an AI roundtable — I was pretty skeptical. Then, starting in January, everything changed for mobile.” It wouldn’t be the first time Mondino got it wrong: back in 2007, he wrote off the iPhone as a passing fad too. “Looking back, it was pretty easy to see how that story was going to end,” he admits.</p><p>To take the measure of just how deep this shift runs, Bitrock gathered nine leading voices for a virtual roundtable: <strong>Android, iOS, Flutter, React Native, and Kotlin Multiplatform developers, joined by two AI specialists.</strong></p><p>This first piece tackles a question that’s simple to ask but harder to answer: <strong>what does AI actually mean today for app developers, and how should they be putting it to use?</strong></p><h3>No Longer a Novelty, but a Daily Habit</h3><p>The first takeaway is that AI has worked its way into everyone’s workflow — only the degree varies. Some rely on it like a chatbot, others let it run as an autonomous agent, and others just use it to quickly get their bearings in unfamiliar code. Marco Gomiero, Principal Android Engineer and Google Developer Expert, is one of the most extreme cases: “I haven’t written a line of code myself in over a year. We use it constantly, at work and on side projects alike. I’ve never been happier.”</p><p>Some teams are already going all-in on fully “agentic” development. Carlo Lucera, Flutter Team Lead at Pivotal Technologies, described a shift his team was living through at that very moment: “We’re moving on from the usual chatbot-in-the-IDE setup toward full agentic development, where our job becomes reviewing pull requests and fixing the code.”</p><h3>The Metaphor Trap: Junior, Senior, or Intern?</h3><p>One of the liveliest debates centered on a gut-level question: does AI act more like a junior teammate, a senior consultant, or an intern who occasionally messes things up? The group’s answer, pretty much in chorus, was that none of these labels quite fit — it all comes down to how the tool is steered.</p><p>Federico Nessi, an iOS developer at Bitrock, comes up with the sharpest analogy: “Think of it as a Formula 1 car — insanely fast if you know how to handle it, but you really need to know what you’re doing.” Stefano Mondino, who says he’s never been comfortable personifying models, would rather picture it as “a keyboard typing at breakneck speed” — one that needs to be set up just right before it writes what you actually have in mind.</p><p>The AI specialists bring the conversation back down to technical reality. Mauro Marinello, Data Scientist and Product Owner at <a href="https://radicalbit.ai/it/">Radicalbit</a> (part of the <a href="https://fortitudegroup.it/">Fortitude Group</a> product portfolio), cuts straight to the point: “A model is nothing more than a token generator. Whether it behaves like a junior or a senior comes down entirely to how well we prompt it.”</p><h3>The Real Engine Is Context</h3><p>And that’s really the crux of it. AI’s value doesn’t come from the model itself, but from the quality and structure of what you feed it. “Picture the model as a box of information,” Marinello explains. “The more organized that information is, the better the output. Type ‘build me a diet app’ and nothing else, and you’ll get junk. But bring in skills, hook it up to an MCP server, and take the time to write a solid system prompt — and everything changes.”</p><p>His colleague Marco Riva, AI engineer at Fortitude Group, shares an example worth pondering for anyone new to these tools: the exact same model, with the exact same setup, can produce wildly different results depending on who’s behind the keyboard. “Hand it to someone with zero development background versus a seasoned expert, and you’ll get completely different outcomes. How much value you get out of the tool depends on how skilled you are.”</p><h3>The Starting Point</h3><p>If there’s a single thread running through this first conversation, it’s this: AI has cranked up speed dramatically, but it hasn’t moved the center of gravity an inch. What still decides the outcome is the skill of whoever’s using it, and how carefully they set the context. It’s an extraordinarily fast keyboard — but somebody still needs to know what to type.</p><p>The next two installments dig deeper: <strong>why agents tend to “gravitate toward” certain frameworks and what that means for tech choices going forward</strong>, and why — even with all this automation — <strong>human expertise and soft skills are still what really sets people apart</strong>.</p><p><strong>Three-part series — </strong><a href="https://www.youtube.com/watch?v=PURSkB7hjt8"><strong>Watch the Roundtable again</strong></a></p><p>Speakers: <a href="https://www.linkedin.com/in/federico-monti-a4b29715b/">Federico Monti</a> (MOLO17), <a href="https://www.linkedin.com/in/stefanomondino/?locale=en">Stefano Mondino</a> (Synesthesia), <a href="https://www.linkedin.com/in/clucera/">Carlo Lucera</a> (Pivotal Technologies), <a href="https://www.linkedin.com/in/marco-gomiero/">Marco Gomiero</a> (Airalo / Google Developer Expert), <a href="https://www.linkedin.com/in/fnessi/">Federico Nessi </a>(Bitrock), Emanuele Maso (Bitrock), <a href="https://www.linkedin.com/in/adallaporta/">Alberto Dallaporta </a>(Novalab), <a href="https://www.linkedin.com/in/mauromariniello/">Mauro Marinello</a> (Fortitude Group), <a href="https://www.linkedin.com/in/marco-riva-098a72205/">Marco Riva </a>(Fortitude Group).</p><p>Moderator: <a href="https://www.linkedin.com/in/samantha-giro-28106ba7/">Samantha Giro</a>, Mobile Manager at Bitrock.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b37e2155bd41" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Payload CMS: When a Standard CMS Just Won’t Cut It]]></title>
            <link>https://medium.com/@BitrockIT/payload-cms-when-a-standard-cms-just-wont-cut-it-cee94207c589?source=rss-5b07e5be1c5e------2</link>
            <guid isPermaLink="false">https://medium.com/p/cee94207c589</guid>
            <category><![CDATA[front-end-development]]></category>
            <category><![CDATA[cms]]></category>
            <category><![CDATA[payload]]></category>
            <dc:creator><![CDATA[Bitrock]]></dc:creator>
            <pubDate>Tue, 07 Jul 2026 09:01:03 GMT</pubDate>
            <atom:updated>2026-07-07T09:01:03.017Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/900/1*Om2KJoKK9z6kp47LF7A3pA.png" /></figure><p>Payload is an open-source headless CMS designed for the Next.js/React stack. Unlike platforms where the schema is configured almost entirely via a graphical interface, in Payload, collections, fields, permissions, and business logic are defined in TypeScript, within the same application repository. It is not just a simple CMS: Payload presents itself as a backend for the modern web featuring content management, authentication, uploads, APIs, a job queue, and an extensible admin panel written in React. Being open-source and self-hosted, your data remains in your Postgres database, without being locked into a closed SaaS.</p><p>For those developing in TypeScript on Next.js, the main advantage is reducing the gap between a “CMS for editors” and an application with complex rules. There is no need to adapt a generic product with fragile plugins: models and behavior live in the code, versioned with Git, testable, and strongly typed.</p><p>In this article, our Front-end developer, Gianluca La Rosa, talks about Payload CMS through a project he realized, in which he was able to test the features of this CMS to build a website and a management application that share the same source of truth.</p><h3>One schema, multiple applications</h3><p>In many projects, the CMS only serves the showcase website. In reality, Payload can become the shared data layer for multiple Next.js apps talking to the same database. In the project from which I extracted these examples, the schema lives in a shared package imported by three distinct applications. The public site consumes the content; the Payload admin runs alongside the frontend; an internal app on a dedicated subdomain handles operations that must not end up in the association’s management panel.</p><p>All apps obtain the same Payload instance with just a few lines:</p><p>import configPromise from ‘@payload-config’</p><p>import { getPayload } from ‘payload’</p><p>export async function getPayloadClient() {</p><p>return getPayload({ config: configPromise })</p><p>}</p><p>No custom REST APIs to maintain in parallel: Server Components and Server Actions directly call payload.find(), payload.update(), and so on. A single schema, a single set of migrations, and shared generated types.</p><h3>Schema in code, not in a panel</h3><p>Collections are defined as TypeScript objects: slugs, fields, hooks, permissions, and admin labels. After every modification, Payload generates a payload-types.ts file aligned with the database, providing autocomplete and type-checking across the entire project.</p><p>For PostgreSQL, it is best to deactivate the automatic “schema push” and rely on versioned migrations:</p><p>db: vercelPostgresAdapter({</p><p>pool: { connectionString: process.env.POSTGRES_URL },</p><p>push: false, // the schema evolves only with payload migrate</p><p>}),</p><p>On a project that lives for months or years, this is important: every change goes through a reviewable SQL migration, applied in CI and production using the exact same command. You do not rely on an implicit sync that “fixes everything” in dev but brings surprises in production.</p><h3>Flexible permissions, beyond admin and editor</h3><p>Many CMS platforms offer only a few fixed roles. Payload exposes access functions on every CRUD operation, with two response modes.</p><p>A binary permission, meaning true or false, across the entire collection:</p><p>export const canCreatePosts: Access = ({ req: { user } }) =&gt; {</p><p>return user?.roles?.includes(‘editor’) ?? false</p><p>}</p><p>Or a query constraint: yes, but only on certain documents. The function returns a where filter, just like in a query:</p><p>export const canReadPosts: Access = ({ req: { user } }) =&gt; {</p><p>if (user?.roles?.includes(‘admin’)) return true</p><p>// Editors only see their own articles</p><p>return { author: { equals: user?.id } }</p><p>}</p><p>Payload does not limit itself to filtering data in memory via Node.js; instead, it translates access constraints directly into SQL WHERE clauses before querying the database. This guarantees maximum performance by leveraging native indexes and eliminates security vulnerabilities at the root: if a user does not have permissions to see a specific record, for the database, that document simply does not exist.</p><p>In the reference project, this pattern is pushed even further: “specialist” roles for different editorial scopes (events, game catalogs, …), differentiated sidebar visibility, and media file permissions tied to a specific scope. By isolating the areas of competence for each user, this made management significantly clearer.</p><h3>Hooks: behavior close to the data</h3><p>Hooks (beforeChange, afterChange, afterRead, …) are where Payload sets itself apart from more content-only CMS platforms. You can execute logic every time a document is read, created, or modified, whether via admin, REST, or the Local API.</p><p>Two typical use cases, both present in the project:</p><ol><li><strong>Derived fields:</strong> values calculated from other fields and the current date, automatically updated instead of being edited manually.</li><li><strong>Side effects:</strong> invalidating the Next.js cache when content is published, synchronizing linked records, or sending notifications.</li></ol><p>A simplified revalidation hook:</p><p>export const revalidatePost: CollectionAfterChangeHook&lt;Post&gt; = ({</p><p>doc,</p><p>req: { context }</p><p>}) =&gt; {</p><p>if (context.disableRevalidate) return doc</p><p>if (doc._status === ‘published’) {</p><p>revalidatePath(`/articoli/${doc.slug}`)</p><p>revalidateTag(‘posts-sitemap’)</p><p>return doc</p><p>}</p><p>}</p><p>In one specific case, instead of filling out a collection’s form to add an element, the user inserts a URL, which triggers a fetch that populates all the fields, saving a massive amount of time.</p><h3>Extensible Admin</h3><p>Being written entirely in React, the admin panel allows you to replace or complement any default element (fields, entire views, sidebar links) by specifying your own component in the configuration file.</p><p>For example, it is possible to define a ui type field that leverages built-in hooks like useAllFormFields to read form data and display a calculated summary instantly as the operator types:</p><p>{</p><p>name: ‘paymentPreview’,</p><p>type: ‘ui’,</p><p>admin: {</p><p>components: {</p><p>Field: ‘@/components/admin/PaymentPreview#PaymentPreview’,</p><p>},</p><p>},</p><p>}</p><p>This approach allows for the creation of dynamic, custom interfaces without resorting to iframes or external plugins. Furthermore, the native Live Preview feature places the frontend preview right next to the editor, allowing you to test various mobile, tablet, and desktop breakpoints during editing.</p><h3>Built-in multilingual support in the CMS</h3><p>Payload handles translations at the field level (localized: true), complete with a default locale and fallback. Supported languages are declared in the config:</p><p>localization: {</p><p>locales: [</p><p>{ code: ‘it’, label: ‘Italiano’ },</p><p>{ code: ‘en’, label: ‘English’ },</p><p>],</p><p>defaultLocale: ‘it’,</p><p>fallback: true,</p><p>},</p><p>Defining translations within the CMS guarantees data model integrity, as the language exists in the schema before the website even displays it. Adding a language requires a controlled migration on Postgres, avoiding the typical misalignment issues found in static JSON files. In this way, the frontend only handles interface translations (buttons, fixed labels) and requests dynamic content from Payload that is already filtered at the source for the correct language.</p><h3>Local API: server-side queries without HTTP</h3><p>The public website does not fetch content via REST from the browser. Server Components query Payload server-side:</p><p>const { docs } = await payload.find({</p><p>collection: ‘events’,</p><p>where: { _status: { equals: ‘published’ } },</p><p>overrideAccess: false,</p><p>locale: ‘it’,</p><p>sort: ‘-date’,</p><p>})</p><p>By utilizing the Local API within Server Components, Payload communicates directly via Node.js code, eliminating HTTP calls and the exposure of secret keys or tokens in the browser. To avoid metadata overhead or the exposure of sensitive hidden data within raw documents, the best practice involves mapping the results into lean DTOs, passing only the properties strictly necessary for rendering to the Client Components.</p><h3>Plugins and infrastructure</h3><p>Official plugins integrate directly into the configuration file, natively extending collections, types, and migrations. Thanks to the integrated Sharp library, image resizing and compression happen on the server during upload. Finally, the native job queue and cron system allow you to handle heavy background tasks using the same application database, without having to configure complex infrastructures.</p><p>It isn’t the WordPress ecosystem, but it covers typical needs; everything else can be extended with hooks and components, using the same language as the rest of the project.</p><h3>Pros and Cons</h3><h3>Why Payload</h3><ul><li><strong>End-to-end TypeScript:</strong> schema, types, frontend, admin.</li><li><strong>Total control:</strong> self-hosted, open-source, no vendor lock-in on data.</li><li><strong>Native Next.js:</strong> admin, API, and site within the same monorepo.</li><li><strong>Extensibility:</strong> hooks, access, and admin UI without artificial limits.</li><li><strong>Performance:</strong> direct DB queries, no mandatory GraphQL.</li></ul><h3>Limitations to consider</h3><ul><li><strong>Initial learning curve:</strong> requires TypeScript/React expertise; it is not “install and forget”, and it is less immediate than WordPress for newcomers.</li><li><strong>Admin for occasional users:</strong> rarely ideal out-of-the-box.</li><li><strong>Ecosystem:</strong> fewer themes and third-party plugins.</li><li><strong>Operations:</strong> database, migrations, and storage are managed by you (unlike Contentful or Sanity).</li><li><strong>Rapid evolution:</strong> major upgrades must be planned.</li></ul><h3>Conclusion</h3><p>Payload is not the tool to choose for a simple blog ready in five minutes. Instead, it becomes the ideal solution when the CMS needs to integrate deeply with the application, acting as a truly flexible backend. It is perfect when you need to share the same schema across multiple apps, manage granular permissions, embed business logic directly into hooks, or customize the admin interface without rebuilding a backoffice from scratch.</p><p>In real and complex contexts — such as an ecosystem combining a multilingual site, structured catalogs, restricted areas, and user management — Payload ensures that the data model and behaviors always remain consistent between the public site, the control panel, and internal applications.</p><p>There is certainly a higher initial investment in terms of development and configuration compared to a traditional CMS or a ready-to-use SaaS platform. However, this effort is amply rewarded over time: you get a robust, clean, and scalable architecture where a single, solid source of truth exists for all project data</p><p>Main Author: <a href="https://www.linkedin.com/in/larosa-gianluca/">Gianluca La Rosa</a>, Front-end developer @ Bitrock</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=cee94207c589" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Predictive Medicine and Real-Time Monitoring: Bitrock’s Approach to Next-Generation Healthcare]]></title>
            <link>https://medium.com/@BitrockIT/predictive-medicine-and-real-time-monitoring-bitrocks-approach-to-next-generation-healthcare-a359c68f0e28?source=rss-5b07e5be1c5e------2</link>
            <guid isPermaLink="false">https://medium.com/p/a359c68f0e28</guid>
            <category><![CDATA[real-time-monitoring]]></category>
            <category><![CDATA[predictive-medicine]]></category>
            <dc:creator><![CDATA[Bitrock]]></dc:creator>
            <pubDate>Tue, 30 Jun 2026 09:15:38 GMT</pubDate>
            <atom:updated>2026-06-30T09:15:38.534Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/900/1*iMxCIKQOe5cHpYJEuEgQvw.png" /></figure><p>Patient management within healthcare facilities is undergoing a fundamental transformation. The maturation of connected medical devices, the development of <strong>AI models applied to biometric signals</strong>, and the consolidation of<strong> event streaming</strong> standards now make a care model practically viable that, until a few years ago, remained confined to research: the patient as a continuously monitored system, capable of generating heterogeneous data streams that, correlated in real time with the clinical history, anticipate the acute event.</p><p>For healthcare facilities and clinical staff, the consequence is an operational paradigm shift. It is no longer about digitising the medical record or centralising ward monitors: the <strong>entire information architecture of clinical activity must be rethought as a structural component of the long-term operational model.</strong></p><p>Against this backdrop, the concept of <strong>proactive protection</strong> takes shape — an architectural approach that distinguishes itself from traditional models through the role played by Artificial Intelligence.</p><p>In conventional systems, devices detect a parameter, the data flows to a ward monitor or departmental system, and only afterwards does the clinician manually correlate signals with the patient’s history. AI, where present, is a layer appended at the end.</p><p>In the logic of proactive protection, by contrast, <strong>AI is the criterion that orients upstream design decisions</strong> — which signals to integrate, how to acquire and correlate them, how to structure model governance — because the<strong> ability to generate predictions and actions in real time</strong> is the starting requirement.In this article we will analyse a concrete scenario: <strong>continuous patient monitoring within complex healthcare facilitie</strong>s. We will start from the problem — clinical latency and healthcare data fragmentation — and work through to the technical architecture Bitrock uses to implement the solution, leveraging the<strong> technologies in the </strong><a href="https://fortitudegroup.it/en"><strong>Fortitude Group</strong></a><strong> product portfolio</strong>. We will then examine the results this model produces on clinical organisation.</p><h3>The Cost of Clinical Latency and Healthcare Data Fragmentation</h3><p>The clinical and operational debt generated by the <strong>absence of an integrated healthcare ecosystem</strong> manifests across three interdependent dimensions:</p><ul><li><strong>Clinical latency</strong>: between the moment a parameter begins to deviate and the moment a significant clinical picture is recognised by staff, minutes or even hours can pass. In areas such as intensive care, post-operative wards, or sub-intensive units, it is precisely within this interval that the difference between a managed event and an acute complication is determined.</li><li><strong>Cognitive load on clinical staff</strong>: ward monitors, laboratory systems, imaging, and EHR each produce their own data stream, across different interfaces and protocols. Correlating these streams remains a manual activity, carried out by the clinician in the limited time available — a mechanism that generates alarm fatigue, spurious alerts, and, in some cases, relevant signals that go unrecognised in time.</li><li><strong>Overall pathway efficiency</strong>: unanticipated complications translate into additional inpatient days, transfers to higher-acuity areas, and resource consumption that, across the patient portfolio managed by a facility, has a structural impact on the spending profile.</li></ul><p>Underlying all this is an information design problem: <strong>healthcare data fragmentation</strong>. The facility is already highly instrumented, but the data streams exist in <strong>separate technological silos</strong>, each with its own evolutionary history and protocols. Correlating these streams is the product of discretionary human intervention — <strong>not scalable, difficult to audit, and ill-suited to a predictive logic that requires cross-referencing signals in real time.</strong>Bitrock addresses this fragmentation by shifting the focus of intervention from application integration to <strong>architectural design</strong>: an <strong>event streaming platform deployed on the healthcare facility’s infrastructure, which acquires data streams at the source, normalises them into a clinically coherent data domain, and makes them available in real time to predictive models and operational agents.</strong></p><h3>Clinical Activity as a Connected System</h3><p>The reference architecture is structured around <strong>four functional layers</strong>, integrated as parts of a unified design. Each layer addresses a specific step in the journey from a single acquired signal to an anticipated clinical intervention.</p><h3>Clinical Stream Ingestion</h3><p>Ward devices generate a constant stream of data acquired via the MQTT protocol and fed into the streaming pipeline. Ingestion is implemented through <a href="https://waterstream.io/"><strong>Waterstream</strong></a>, the MQTT broker from the Fortitude portfolio natively integrated in Apache Kafka: a choice that <strong>eliminates the translation layer between MQTT and Kafka</strong> and reduces latency between the device and the analytics system to the technically sustainable minimum. The platform is deployed on the healthcare facility’s infrastructure — a necessary condition both for compliance with sensitive data processing regulations and for the operational sustainability of the solution.</p><h3>Low-Latency Multimodal Correlation</h3><p>Individual signals are transformed into risk patterns. The <strong>stream processing middleware</strong>, built with <strong>Apache Flink</strong> in alignment with the event streaming stack supported by Waterstream and Kafka, cross-references biometric data in real time with contextual information: the patient’s clinical history and EHR, ongoing therapies and administrations, laboratory results, and ward reports. The Waterstream–Kafka–Flink pipeline is the enabling factor behind the system’s <strong>low end-to-end latency</strong>: from the signal detected by the device to the generation of a clinician-usable prediction, every step is executed in streaming, with no intermediate batch accumulation phases. It is at this layer that predictive logic delivers its real value: the identification of risk combinations <strong>that cannot be reduced to reading a single monito</strong>r — micro-deviations that, read together, anticipate clinically significant scenarios.</p><h3>Governed Predictive Inference</h3><p>The third layer is that of governed predictive inference, where the patterns identified at the previous level are <strong>transformed into operational predictions</strong>. AI models recognise silent clinical deterioration: combinations of micro-anomalies that, taken individually, remain below the alarm thresholds of traditional monitoring systems, but that read over time anticipate risk scenarios.</p><p>In Bitrock projects, this layer is orchestrated through the <a href="https://radicalbit.ai/"><strong>Radicalbit</strong></a> <strong>MLOps platform,</strong> part of the Fortitude product portfolio, which internalises the requirements of the healthcare sector as native capabilities: audit trail, end-to-end tracing, personal data masking, cost control, and inference observability are b<strong>ehaviours of the gateway, not features to be implemented in the application-layer code that consumes the predictions.</strong> In a domain where every component is subject to validation and the traceability of each individual prediction is a condition of operation, this property makes a decisive difference.</p><p>Added to this is the <strong>decoupling between the application and the underlying AI model</strong>. The model can be replaced, retrained, or updated without changes to the clinical platform’s code. Over the typical lifecycle of a hospital platform, this property makes it possible to evolve technological choices at the pace of the clinical-scientific state of the art, without every model update becoming a refactoring project — or, more critically, a new end-to-end validation cycle.</p><h3>Orchestrated Response</h3><p>In this final layer, <strong>the prediction is translated into coordinated clinical action</strong>. When predictive risk thresholds are exceeded, the system activates a chain of agents operating across three dimensions.</p><p>At the <strong>clinical level</strong>, the professional does not receive a simple alarm but a pre-compiled diagnostic picture, delivered through a high-performance front-end following a Digital Delivery logic: detected anomaly, correlation with ongoing therapies, and suggested indications drawn from RAG models backed by the relevant literature and the facility’s internal clinical protocols.</p><p>At the <strong>operational level</strong>, automated internal escalation protocols are activated: notification to the rapid response team, request for transfer to a higher care intensity level, and allocation of the instrumental and personnel resources needed to manage the event.</p><p>At the <strong>relational level</strong>, a conversational assistant based on a Small Language Model can support interaction with the patient for qualitative information gathering or for initial contact while awaiting the clinician. <strong>Irreversible clinical decisions remain with the healthcare professional; but the system compresses the time needed to build the context on which those decisions are made.</strong></p><h3>Results: From Reactive to Predictive and Adaptive Care</h3><p>The strategic objective of a proactive protection architecture is the transformation of the care model: the shift from a reactive approach to a <strong>predictive and adaptive model, in which clinical resources are concentrated where and when the data justifies it.</strong></p><p>Operational results are structured across three dimensions:</p><ul><li><strong>Reduction in the time to recognise clinical deterioration</strong>, achieved through automatic signal correlation and early identification of risk patterns.</li><li><strong>Optimisation of clinical workload</strong>, with reduced alarm fatigue, intervention focused on genuinely high-priority cases, and a pre-structured information picture to support decision-making.</li><li><strong>Clinical data security and governance</strong>, guaranteed by an infrastructure compliant with the GDPR and the European Health Data Space, in which the traceability of each individual prediction is a native platform property.</li></ul><p>In terms of operational costs, transitioning to this model makes it possible to anticipate interventions that, in a reactive scenario, would have generated additional inpatient days, transfers to higher-acuity areas, and avoidable complications. For entities managing large healthcare facilities, this optimisation manifests as a <strong>structural shift in the spending profile</strong>, accompanied by a measurable improvement in clinical outcomes.</p><h3>Conclusion</h3><p>Proactive protection is not a new product category: it is a<strong> different way of conceiving clinical activity within healthcare facilities</strong>. It involves integrating streams that have historically lived in silos, introducing predictive logic operating in real time, and building a response chain that combines orchestrated automation with structured clinical decision-making.</p><p>In the short term, this model will determine healthcare operators’ ability to manage increasingly complex and growing caseloads within a regulatory framework set to demand ever more stringent standards.</p><p>The value of a partner like <strong>Bitrock</strong>, integrated within the <strong>Fortitude Group</strong>‘s vision, lies in the ability to approach this transition as a single coherent design: from strategy to architecture, from technology components to delivery, through to the operational model that accompanies the platform in production — in a domain where the reliability of each individual component is a condition of operation.</p><p>Vertical technological depth, an Agile method applied to enterprise complexity, and a product portfolio designed for the most critical points in the architecture combine into a set of levers that allows the client to evolve their care model without accumulating technical debt or irreversible dependencies.</p><p>This is the difference between <strong>adopting a healthcare technology and building a clinical information infrastructure capable of generating lasting value over time.</strong></p><p><a href="https://bitrock.it/contacts">Contact us</a> to explore with our experts how a proactive protection architecture can fit your organisation’s specific context and challenges.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a359c68f0e28" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Coding Agents and Software Development: Understand Before You Act]]></title>
            <link>https://medium.com/@BitrockIT/coding-agents-and-software-development-understand-before-you-act-71f071b0edb7?source=rss-5b07e5be1c5e------2</link>
            <guid isPermaLink="false">https://medium.com/p/71f071b0edb7</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ai-coding-agent]]></category>
            <category><![CDATA[ai-software-developer]]></category>
            <dc:creator><![CDATA[Bitrock]]></dc:creator>
            <pubDate>Wed, 24 Jun 2026 08:08:04 GMT</pubDate>
            <atom:updated>2026-06-24T08:08:04.076Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/900/1*WGnjpQs5MykXgSl_9Do1nA.png" /></figure><p>Although it might seem strange to most, the first launch of ChatGPT — which sparked what we now call the AI boom — happened less than four years ago. Since then, most of us have witnessed an incredible change in the way we work, across a wide variety of jobs.</p><p>Software development was one of the most affected fields, and the reactions were understandably mixed. The first wave of concern was existential: would developers be made redundant? I do not believe this is the case. As in all major revolutions in history humans have never been made redundant, they just adapted, so developers will adapt too. But in all adaptation processes some old ways of doing things will disappear and new ways will emerge. This process of change can be painful, because it comes with a form of resistance, an inertia around changing the way you work and the way you think.</p><p>To overcome that resistance, I believe that it’s important to have an abstract idea of how an agent actually works, because it can help you understand what it is and isn’t, where it shines and where it doesn’t, allowing you to take full advantage of this amazing technology. In this post, I want to share my mental picture of agents and the practical habits that follow from it. I want to make clear that this is what worked for me, it may not be the only way or the right way to think about it, but having a clear abstraction in mind definitely made my work easier.</p><p>The mental picture draws on <a href="https://www.youtube.com/watch?v=LCEmiRjPEtQ">ideas popularized in 2025 by Andrej Karpathy</a>.</p><h3>A Mental Model for Agentic Systems</h3><p>We are probably all familiar with the main hardware components of a computer: At the core of a computer is the processing unit, i. e. the CPU. The CPU is the place where computation actually happens — where bits of information are processed. All the other components are builded around it: the RAM, the hard disk, the keyboard, the display. These components are part of the hardware, which is the substrate where computation lives. But hardware alone does not specify what to compute, for that you need software.</p><p>On the software side, the Operating System (OS) orchestrates the flow of bits from external sources (the hard disk, keyboard inputs, and so on) through the CPU, and back to the display, or any other output. But it still doesn’t specify how this information is processed. This last step — the critical step for the resolution of the specific problem — is specified in a program: a set of instructions that defines how bits of information must be processed for a specific task.</p><p>At this point, note if we stop thinking about “bits” of information and we start to think about “tokens” of information — i. e. written text, you might see that an agentic system operates under a very similar conceptual scheme.</p><p>The OS is analogous to the agent’s harness. You can think about the harness as the TypeScript or Python code that manages the sessions and orchestrates tool execution. To go farther, you can think about the model’s context window as the RAM, and operations like context compression, which are part of the agent’s harness, are analogous to an OS freeing up RAM when the maximum capacity is close. But it is in the program, subprograms, runtime and CPU parallels that the analogy gets more interesting.</p><h3>The Program, the Runtime, and the Natural-Language CPU</h3><p>In a traditional computer, the program is a fixed set of instructions written before execution. These instructions might also imply that other subprograms are called during the execution in order to perform specific operations. The runtime is the live state that accumulates while the program runs.</p><p>In an agentic system, the program is the initial task — the first user message, or the task assigned by a higher-level agent to a subagent. A subprogram maps to a skill: a pre-written set of instructions that can be invoked during the execution of the main program for a specific operation. The runtime is the session — the input context that accumulates in the context window as the program (the task) is executed.</p><p>But what about the CPU? In a traditional computer, the CPU executes binary instructions deterministically. The LLM occupies the CPU position — but it is not a CPU. It is a natural language processor. This difference is important: the CPU speaks binary, the LLM processes natural language — the same medium the human uses. This makes the boundary between human and machine far more permeable: the “brain” of the machine speaks human now, thus the human can contribute to the runtime directly, far more easily than with a CPU, without a formal specification standing between them and the execution.</p><p>Moreover, the shared language implies that the program can be steered during execution. Each message shapes the context, and therefore the output, for every subsequent turn. The human is no longer outside the execution — they are a co-processor of the runtime, alongside the LLM.</p><p>From this mental picture, I drew three practical consequences.</p><h3>Three Habits, One Principle</h3><h3>1. Start by talking, not doing</h3><p>Working with a coding agent is not configuring a machine. It is beginning a collaborative session with a peer processor. The opening turns establish the context that shapes the probability distribution for everything that follows: how the problem is framed, what constraints are clarified, what assumptions are made explicit.</p><p>As a rule of thumb, the first move in a session should be a question, not an instruction. A few minutes of exploration before the first command often produces better outcomes. In this regard, you might want to check out <a href="https://github.com/mattpocock/skills/tree/main/skills/engineering/grill-with-docs">grill-with-docs</a>, a Claude Code skill built by Matt Pocock which I use a lot before starting building code.</p><h3>2. Beware of the tangent</h3><p>Exploration is natural and often productive, but can also be bad if it goes too far, that is if it loses focus on the main task. Asking a related question or considering a different approach can be useful, but sometimes the conversation can go on a tangent, which in our analogy means that the system starts running a different unrelated main program in the same runtime. A tangent in a session does not disappear. Even if you return back to the main task, its content stays in the input context, consuming tokens and confusing the agent.</p><p>The fix is a fresh session. When you recognise the tangent, you run something like <a href="https://github.com/mattpocock/skills/tree/main/skills/productivity/handoff">handoff</a>, another skill by Matt Pocock. It compresses the current session into a focused summary that a new session can pick up from immediately. In our analogy, it’s a self-contained subprogram that will create the main program to be run in a new runtime to develop the idea in the tangent, while the original runtime can continue focused on the original program.</p><h3>3. You can outsource your thinking, but not your understanding</h3><p>The specs-to-code paradigm — write a full specification, hand it over, let the agent build everything autonomously — tries to restore the original boundary between the human and the machine: human as author outside, agent as executor inside. It may work, but in any case you lose the steering advantage. The human is no longer a co-processor — they are waiting for output, which must then be reviewed and understood by the human in any case, and all at once.</p><p>The alternative is incremental steps: break the problem into small units, complete one at a time, review each before moving to the next. Remember that the LLM can go on tangents just as the human can. Incremental review is the only way to catch them before their content accumulates in the input context and compounds. If you understand each step before proceeding you are kept inside the execution.</p><p>“<a href="https://x.com/karpathy/status/2049907410303865030">You can outsource your thinking, but not your understanding</a>” describes beautifully the cause of what can go wrong when a coding agent is treated too much as an autonomous contractor rather than a collaborative runtime.</p><h3>Conclusion</h3><p>An agentic system architecture is similar to a traditional computer architecture, but with a CPU that speaks human language. The LLM is the CPU, the session is the runtime, and you are a co-processor. You can program the agent in plain English and you can steer the runtime acting directly on the CPU. Three practical habits follow naturally from this view: start by establishing context before issuing commands, recognise tangents before they compound, and stay inside the execution through incremental review. Remember that you can outsource your thinking, but not your understanding.</p><p>Main Author: <a href="https://www.linkedin.com/in/giovanni-vacanti/">Giovanni Vacanti</a>, Data Scientist @ Bitrock</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=71f071b0edb7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Predictive Resilience for Strategic Infrastructure: The Next Generation of Smart Cities]]></title>
            <link>https://medium.com/@BitrockIT/predictive-resilience-for-strategic-infrastructure-the-next-generation-of-smart-cities-b612e65461fb?source=rss-5b07e5be1c5e------2</link>
            <guid isPermaLink="false">https://medium.com/p/b612e65461fb</guid>
            <category><![CDATA[iot]]></category>
            <category><![CDATA[smart-city-solutions]]></category>
            <category><![CDATA[predictive-analytics]]></category>
            <dc:creator><![CDATA[Bitrock]]></dc:creator>
            <pubDate>Tue, 09 Jun 2026 09:01:02 GMT</pubDate>
            <atom:updated>2026-06-09T14:09:24.371Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/900/1*d6FwnuyZZ2mTQr54IO2SJw.png" /></figure><p>The management of strategic urban infrastructure — bridges, viaducts, tunnels, mobility hubs — is undergoing a fundamental shift. The <strong>maturity of IoT technologies</strong>, the development of <strong>AI models applied to structural monitoring</strong>, the spread of <strong>edge computing</strong>, and the consolidation of <strong>event streaming</strong> standards now make it possible to deploy a management model that, only a few years ago, was still confined to research: infrastructure as a living system, capable of acquiring heterogeneous data streams in real time, correlating them, and generating preventive risk forecasts.</p><p>For the companies that manage these assets, the consequence is a shift in operating paradigm. It is no longer a matter of modernizing an existing monitoring system: it is necessary to <strong>rethink the asset’s information architecture as a structural component of the long-term operating model</strong>. The European regulatory framework, moreover, has made this evolution an explicitly required line of development for infrastructure classified as critical.</p><p>It is precisely in this scenario that the concept of <strong>predictive resilience</strong> takes shape — an architectural approach that differs from traditional models in the role played by Artificial Intelligence.</p><p>In conventional systems, sensors collect data, the data feeds a monitoring system, and at a later stage an analytical model is applied to the already-structured data to extract insights: AI is a layer added at the end of the chain.</p><p>In the logic of predictive resilience applied to smart cities, by contrast, <strong>AI is the criterion that drives upstream design choices</strong> — which sensors to install, how to acquire the streams, where to process them, how to structure model governance — because the <strong>capacity to generate forecasts and actions in real time</strong> is the starting requirement.</p><p>In this article we will analyze a concrete use case: the <strong>predictive management of a large strategic urban infrastructure within a smart city ecosystem</strong>. We will start from the problem — the cost of congestion and data fragmentation — and move through to the technical architecture with which <strong>Bitrock</strong> implements the solution, leveraging the <strong>technologies of the Fortitude Group product portfolio</strong> that enable its critical components. Finally, we will analyze the results that the model delivers in the operational management of the asset.</p><h3>The Cost of Congestion and Data Fragmentation</h3><p>The analysis starts by quantifying the cost of inertia — namely, the technical and operational debt generated by the <strong>absence of an integrated urban ecosystem</strong>. This burden manifests itself through three interdependent dimensions that act as systemic bottlenecks, amplifying the degradation of urban performance.</p><p>First, infrastructural inefficiency is not limited to mere congestion: it translates into accelerated <strong>mechanical and structural stress</strong>. The absence of optimized flows degrades physical assets and extends travel latency, directly impacting the scalability of urban quality of life. To this is added the <strong>compromise of safety parameters</strong>, where network saturation reduces operational margins and increases the fail rate of emergency interventions, making the road system inherently less resilient. Finally, the sustainability gap highlights how emission peaks at congestion points act as an <strong>environmental leak</strong>: these pollution hotspots undermine the energy optimization processes implemented in other segments, preventing the achievement of a real decarbonization balance at the level of the entire ecosystem.</p><p>Upstream of these three dimensions lies a <strong>structural problem of information design</strong>: <strong>data fragmentation</strong>. In modern urban contexts, critical assets are already significantly instrumented — sensors, cameras, weather stations, weigh-in-motion systems, fleet telemetry — but the resulting information flows live in <strong>separate technological silos</strong>, each with its own evolutionary history and its own protocols.</p><p>The correlation between these flows is the product of discretionary human intervention. A mechanism that is not scalable, hard to audit and slow — overall <strong>unsuited to a predictive logic that requires the cross-referencing of signals in real time.</strong></p><p>Bitrock addresses this fragmentation by shifting the level of intervention from application integration to <strong>architectural design</strong>, designing an <strong>event streaming platform that acquires the flows directly at the source, normalizes them in a shared data domain, and makes them available in real time to predictive models and operational agents.</strong></p><p>It is precisely in this choice that the <strong>AI-ready Data Ecosystem approach of the Fortitude Group</strong> takes concrete form.</p><h3>Infrastructure as a Dynamic Entity</h3><p>Bitrock approaches platform design around a reference architecture organized in <strong>four functional layers</strong>, integrated as parts of a single, unified design.</p><h3>Ingestion and Edge Intelligence</h3><p>Large infrastructure assets are equipped with thousands of <strong>IoT devices that generate data streams acquired via the MQTT protocol</strong> and brought into a <strong>streaming environment</strong> where processing happens close to the asset. In Bitrock’s reference architecture, ingestion is delivered through <a href="https://waterstream.io/"><strong>Waterstream</strong></a>, the Fortitude portfolio’s MQTT broker natively integrated into an Apache Kafka environment: a choice that eliminates the translation layer between MQTT and Kafka. The choice of edge processing, in turn, is not a latency optimization: it is a design requirement to guarantee operational continuity even in the presence of degraded connectivity to the centralized cloud.</p><h3>Multimodal Correlation</h3><p>The single signal is transformed into a risk pattern. The <strong>stream processing middleware</strong>, typically built on <strong>Apache Flink</strong> in line with the event streaming stack supported by Waterstream, <strong>cross-references the data</strong> to generate risk forecasts. A recurring architectural example is the detection of the simultaneous transit of multiple exceptional loads, obtained by combining camera analytics and weigh-in-motion plate data. It is at this layer that <strong>predictive logic produces real value</strong>: the identification of risk combinations that cannot be inferred from the reading of any single sensor.</p><h3>Governed Predictive Inference</h3><p>The third layer is that of governed predictive inference, where the <strong>risk patterns</strong> identified at the previous layer are <strong>transformed into operational forecasts</strong>. AI models recognize silent structural degradation: combinations of micro-anomalies that, taken individually, remain below the alarm thresholds of a traditional monitoring system but that, read together and over time, anticipate significant risk scenarios.</p><p>In Bitrock’s projects, this layer is orchestrated through <a href="https://radicalbit.ai/ai-gateway/"><strong>Radicalbit’s AI Gateway</strong></a>, part of the Fortitude product portfolio, which internalizes all of these requirements as native platform capabilities: audit trail, end-to-end tracing, personal data masking, cost control and inference observability are gateway behaviors, not features to be implemented in the code of the application that consumes the predictions.</p><p>Added to this is a strategically relevant property: the decoupling between the application and the underlying AI model. The model can be replaced, retrained or updated without any change to the application code. Over the lifecycle horizons of an infrastructure asset, this property is what makes it possible to evolve technological choices at the pace of the state of the art, without every model upgrade becoming a refactoring project.</p><h3>Orchestrated Response</h3><p>In this final layer, <strong>the forecast translates into coordinated action</strong>. When predictive safety thresholds are exceeded, the system activates a chain of intelligent agents articulated along three lines. On the <strong>mobility front</strong>, the agents communicate directly with road signs and intelligent traffic lights to deliver an adaptive traffic block, diverting vehicle flow before the risk becomes critical.</p><p>On the <strong>operational front</strong>, automated emergency protocols are activated: immediate dispatch of drones for visual inspection of joints flagged as critical, and priority notification to maintenance teams, accompanied by a technical report already structured by the AI in a Digital Delivery logic.</p><p>On the <strong>decision-making front</strong>, the organization and the competent authority receive, through a high-performance front-end, the complete picture: the anomaly detected, the impact simulation, the intervention options suggested by RAG models grounded in the asset’s technical manuals. Decisions remain with the human operator; but <strong>the system structurally reduces the time required to build the context on which those decisions are made.</strong></p><p><strong>Architectural LayerFunction and Design RationaleIngestion and Edge Intelligence</strong>Real-time acquisition of IoT streams (vibration, deformation, corrosion, environmental parameters, computer vision, load telemetry) via MQTT, with processing close to the asset. Guarantees operational continuity even with degraded connectivity to the centralized cloud.<strong>Multimodal Correlation</strong>Time-windowed stream processing across structural data, weather conditions, traffic telemetry and static/dynamic loads. Transforms isolated signals into risk patterns that cannot be derived from single-sensor readings.<strong>Governed Predictive Inference</strong>AI models for the identification of silent structural degradation, operated in production with full traceability, audit trail and data governance. A non-negotiable requirement for assets that fall within the perimeter of strategic infrastructure.<strong>Orchestrated Response</strong>Intelligent agents that coordinate adaptive traffic blocking, automated emergency protocols and decision support for management. Irreversible decisions remain with the human operator.</p><h3>Results: From Reactive to Predictive and Adaptive Maintenance</h3><p>The strategic objective of a predictive resilience architecture is the transformation of the maintenance model: the shift <strong>from a reactive approach to a predictive and adaptive model, in which resources are concentrated where and when the data justify it.</strong> This is the ground on which the actual return on investment is measured.</p><p>Operational results are organized along three lines:</p><ul><li><strong>Reduction in travel times</strong>, achieved through the intelligent distribution of vehicle loads across the entire urban network.</li><li><strong>Environmental sustainability</strong>, with a reduction in CO₂ emissions linked to smoother traffic flow and the elimination of queues.</li><li><strong>Safety and risk management</strong>: the reduction in response times for critical vehicles and the structural improvement in emergency response, achieved through the combination of dynamic routing, anticipation of structural anomalies and a structured information framework available to the decision-maker.</li></ul><p>In terms of <strong>impact on management costs</strong>, the transition to this new predictive model makes it possible to <strong>avoid redundant scheduled interventions</strong> and to <strong>bring forward interventions</strong> on assets that, under a reactive model, would have generated significantly higher downtime costs.</p><p>For companies that manage extensive infrastructure portfolios, this optimization manifests itself as a meaningful structural shift in the cost profile.</p><h3>Conclusion</h3><p>Predictive resilience is not a new product category: it is a different way of conceiving the <strong>management of strategic infrastructure</strong>. It involves the integration of flows that have historically lived in silos, the introduction of predictive logic operated in real time, and the construction of a response chain that combines orchestrated automation with structured human decision-making.</p><p>In the near term, it is the model that will determine the capacity of organizations to manage complex assets in an <strong>evolving urban context</strong> and within a <strong>regulatory framework</strong> set to demand increasingly stringent standards.</p><p>The value of a partner like <strong>Bitrock</strong>, integrated within the vision of the <strong>Fortitude Group</strong>, lies in the ability to address this transition as a single unified design: from strategy to architecture, from technology components to delivery, all the way through to the operating model that supports the platform in production.</p><p>Deep vertical technological expertise, an Agile method applied to enterprise complexity, and a portfolio of products designed for the most critical points of the architecture combine to form a set of levers that allow the client to evolve its management model without accumulating technical debt or irreversible dependencies.</p><p>It is the difference between adopting a technology and building an <strong>information infrastructure capable of generating stable value over time.</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b612e65461fb" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>