<?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 rizahorasan on Medium]]></title>
        <description><![CDATA[Stories by rizahorasan on Medium]]></description>
        <link>https://medium.com/@rizahorasan?source=rss-84be84f2acd4------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*dmbNkD5D-u45r44go_cf0g.png</url>
            <title>Stories by rizahorasan on Medium</title>
            <link>https://medium.com/@rizahorasan?source=rss-84be84f2acd4------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 10:30:08 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@rizahorasan/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[Closing the AI Adoption Gap: Building Enterprise AI Applications with Spring Boot and LangChain4j]]></title>
            <link>https://rizahorasan.medium.com/closing-the-ai-adoption-building-enterprise-ai-applications-with-spring-boot-and-langchain4j-5118555161c2?source=rss-84be84f2acd4------2</link>
            <guid isPermaLink="false">https://medium.com/p/5118555161c2</guid>
            <category><![CDATA[spring-boot]]></category>
            <category><![CDATA[ai-adoption]]></category>
            <category><![CDATA[enterprise-ai-solutions]]></category>
            <category><![CDATA[langchain4j]]></category>
            <category><![CDATA[langfuse]]></category>
            <dc:creator><![CDATA[rizahorasan]]></dc:creator>
            <pubDate>Wed, 20 May 2026 10:11:00 GMT</pubDate>
            <atom:updated>2026-05-20T14:28:17.078Z</atom:updated>
            <content:encoded><![CDATA[<p>In this article, I will walk through a Spring Boot application that demonstrates modern AI design patterns using LangChain4j.</p><p>In regulated industries, the conversation around LLM applications immediately gravitates toward Python notebooks, experimental frameworks. While Python absolutely dominates AI research and rapid prototyping, production enterprise systems operate under a very different set of constraints. The platforms in regulated environments require:</p><ul><li>operational stability</li><li>observability</li><li>governance</li><li>distributed tracing</li><li>maintainability</li><li>strong typing</li><li>security controls</li><li>integration with existing enterprise ecosystems</li></ul><p>These are areas where the Java ecosystem has decades of maturity.</p><p>Today, Spring Boot, Spring AI and LangChain4j provide a highly capable and production-ready stack for building enterprise-grade LLM applications.</p><p>I will go over some of the fundamental enterprise software engineering problems.</p><ul><li>How do we coordinate multiple LLM tasks?</li><li>How do we route requests dynamically?</li><li>How do we enforce structured outputs?</li><li>How do we trace model executions?</li><li>How do we integrate AI into existing enterprise services?</li><li>How do we monitor token usage and failures?</li><li>How do we preserve auditability?</li></ul><p>I think, Spring Boot and LangChain4j fit naturally for helping these requirements.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4NSLEankY4AwZmbtlcSDqg.png" /><figcaption>Spring Boot, Langchain4j and Langfuse</figcaption></figure><h3><strong>What Is LangChain4j?</strong></h3><p><a href="https://github.com/langchain4j/langchain4j">LangChain4j</a> is a Java framework for building LLM-powered applications. Conceptually, it brings many of the orchestration capabilities popularized in Python ecosystems into the JVM world, while integrating naturally with:</p><p><em>Spring Boot, dependency injection, enterprise observability, typed Java models, existing backend architectures</em></p><p>For enterprise teams already invested in Java, this dramatically lowers adoption friction. Instead of building disconnected AI prototypes, teams can integrate LLM workflows directly into existing enterprise systems.</p><h3><strong>Architecture Overview</strong></h3><p>The application is intentionally structured around enterprise-friendly architectural boundaries.</p><p>The system contains three primary layers:</p><ul><li>Orchestrator Services</li><li>Declarative Workflows</li><li>Functional Nodes</li></ul><p>A shared AgenticScope carries state across workflow execution.</p><p>The architecture enables:</p><ul><li>modular workflows</li><li>reusable nodes</li><li>strongly separated responsibilities</li><li>dynamic orchestration</li><li>parallel execution</li><li>observable execution chains</li></ul><p>At a high level, the flow looks like this:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*epmLixP99rAgKygwHvPynQ.png" /><figcaption>Client → Orchestration Service → Workflow → Agentic Nodes → LLM Models</figcaption></figure><p>The orchestration services are responsible for:</p><ul><li>starting telemetry spans</li><li>managing workflow invocation</li><li>propagating execution context</li><li>coordinating workflow execution</li></ul><p>The workflow layer defines execution topology using LangChain4j’s fluent APIs. This includes sequential execution, conditional routing, fan-out / fan-in parallelization.</p><h3>AI Design Patterns Implemented</h3><p>One of the primary goals of this project is demonstrating practical AI design patterns that can be reused across enterprise applications.</p><ol><li><strong>Prompt Chaining</strong></li></ol><p>Prompt chaining executes multiple prompts sequentially, where the output of one step becomes the input of another. This pattern is especially useful in banking scenarios such as:</p><ul><li>fraud investigation workflows</li><li>multi-step document analysis</li><li>incident triage</li><li>compliance processing</li></ul><p>In the implementation, multiple log readers analyze different information sources step-by-step before producing a consolidated result. The workflow remains modular because each node owns a single responsibility.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/914/1*U6lwVnqoaj-E6AkkPbihOg.png" /><figcaption><strong>Prompt Chaining (via Langfuse)</strong></figcaption></figure><p>2. <strong>Parallelization</strong></p><p>Instead of waiting for sequential model calls, multiple tasks execute concurrently. The implementation uses a classic Fan-out / Fan-in model: multiple nodes execute simultaneously, results are aggregated afterward and latency is significantly reduced.</p><p>This becomes extremely valuable for enterprise systems processing:</p><ul><li>operational logs</li><li>compliance records</li><li>transaction summaries</li><li>customer support analysis</li></ul><p>One important detail is that observability context propagation is preserved across threads. This is often overlooked in AI demos but becomes essential in production systems. Without proper trace propagation, distributed workflows become operationally opaque.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/850/1*g_nW6Jyr1vMox9UpORGbrg.png" /><figcaption><strong>Parallelization (via Langfuse)</strong></figcaption></figure><p>3. <strong>Routing</strong></p><p>Routing enables workflows to dynamically redirect execution based on LLM decisions. The application first classifies a request and then conditionally routes execution to the correct processing node.</p><p>This mirrors real enterprise workflows such as:</p><ul><li>customer request routing</li><li>incident categorization</li><li>document processing pipelines</li><li>fraud escalation flows</li></ul><p>The key advantage is composability. Instead of creating one massive AI agent responsible for everything, smaller specialized agents collaborate through orchestration. This is generally a more maintainable and governable approach for enterprise environments.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*3N7dlEQBGhS2QpeGcp2nEQ.png" /><figcaption>Routing for email logs (via Langfuse)</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gTg_0B6x8uEMeZ80V99wWQ.png" /><figcaption>Routing for database logs (via Langfuse)</figcaption></figure><p>4. <strong>Structured Outputs</strong></p><p>Structured outputs are arguably one of the most important enterprise AI capabilities. In consumer demos, free-text responses are acceptable. In enterprise systems, they are not. Regulated systems require deterministic outputs that downstream services can validate and process safely.</p><p>For example, a natural language request can be mapped into: category enums, confidence scores, and validated response structures.</p><p>Instead of parsing unreliable free-form text, downstream systems receive strongly typed objects. This dramatically improves: reliability, auditability, <br>integration safety, and automation readiness.</p><p>For regulated industries, this capability is critical.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*3N7dlEQBGhS2QpeGcp2nEQ.png" /><figcaption>Structured Output for the input intension</figcaption></figure><h3>Final Thoughts</h3><p>I think, the future of enterprise AI will not be defined solely by model quality. It will also be defined by:</p><ul><li>orchestration</li><li>governance</li><li>observability</li><li>maintainability</li><li>integration capability</li><li>operational excellence</li></ul><p>Perhaps the most important advantage of this approach is how naturally AI capabilities integrate into existing Java enterprise ecosystems. Spring Boot and LangChain4j do not force organizations to reinvent their architecture or abandon established engineering practices. Instead, LLM workflows become another composable part of the existing platform (alongside REST APIs, messaging systems, security layers, observability stacks, and business services).</p><p>For organizations that already operate large JVM-based systems, this dramatically reduces adoption friction. Existing Spring Boot developers can build, deploy, monitor, and maintain AI-powered services using familiar tools and operational models rather than introducing an entirely separate AI engineering stack.</p><p>In many ways, this is where LangChain4j becomes especially compelling: it allows enterprises to extend their current architecture into the AI domain instead of replacing it.</p><p>The repo is avaliable on <a href="https://github.com/horasan/ai-patterns-with-langchain4j">github/horasan</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5118555161c2" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Closing the AI Adoption Gap with Low-Risk, High ROI Projects (Part 3): The “Search and Repeat”…]]></title>
            <link>https://rizahorasan.medium.com/closing-the-ai-adoption-gap-with-low-risk-high-roi-projects-part-3-the-search-and-repeat-5f07a2ad776f?source=rss-84be84f2acd4------2</link>
            <guid isPermaLink="false">https://medium.com/p/5f07a2ad776f</guid>
            <category><![CDATA[generative-ai-in-banking]]></category>
            <category><![CDATA[roi]]></category>
            <category><![CDATA[data-privacy]]></category>
            <category><![CDATA[generative-ai-in-finance]]></category>
            <category><![CDATA[fintech]]></category>
            <dc:creator><![CDATA[rizahorasan]]></dc:creator>
            <pubDate>Thu, 23 Apr 2026 18:26:33 GMT</pubDate>
            <atom:updated>2026-04-23T18:26:33.943Z</atom:updated>
            <content:encoded><![CDATA[<h3>Closing the AI Adoption Gap with Low-Risk, High ROI Projects (Part 3): The “Search and Repeat” Pattern</h3><p>In <a href="https://medium.com/@rizahorasan/closing-the-ai-adoption-gap-with-a-low-risk-high-roi-projects-in-finance-p1-51846ad2957f">Part 1</a>, I explored how “Low-Risk, High-ROI” project can help closing the AI Adoption Gap. Today, I am looking at a common <strong>“Search and Repeat” </strong>pattern without breaking the <strong>Data Privacy </strong>rules in finance.</p><p>The <strong>Search and Repeat</strong> pattern: a workflow that turns natural language into immediate action without compromising sensitive information.</p><p><strong>The problem: </strong>In finance, users spend a significant amount of time performing “repetitive variations.” Usually, that looks like:</p><ol><li>Navigating to a transaction history screen.</li><li>Fiddling with 5+ filters (Date, Customer ID, Product Type).</li><li>Finding a previous deal to use as a template.</li><li>Manually re-typing that data into a new transaction form.</li></ol><p>This is a “hidden tax” on productivity and a breeding ground for manual entry errors.</p><p><strong>The Solution</strong>: Intelligent Retrieval + “Execute Same”</p><p>Instead of a complex UI with filters, the user types:</p><blockquote>“Show me my last 3 SPOT transactions for Customer C001”</blockquote><p>The AI understands the intent, returns the filter parameters and the related service name for querying the data.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/776/1*jsISIB5x3s-mnYT1y9LODQ.png" /><figcaption>Intelligent Retrieval + “Execute Same”</figcaption></figure><p>UI also provides an <strong>“Execute Same”</strong> button. One click, and the user is redirected to a pre-filled transaction screen, according to the <em>Human in the Loop</em> pattern.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/761/1*DfQEoLxRNNJrbW8aolEwBQ.png" /><figcaption>Prefilled transaction screen</figcaption></figure><h3>Why This is Low-Risk: The Decoupled Architecture</h3><p>The primary reason AI projects stall in finance is the <em>Privacy Wall</em>. Sending a customer’s entire transaction history to a public LLM is generally not excepted.</p><p>The pattern bypasses this using following artifacts:</p><ul><li><strong>The LLM as a Translator:</strong> LLM is used to extract <em>metadata</em> only. The LLM identifies that the user wants 3 records of type SPOT,and the related backend service name. It never sees the actual results of the query.</li><li><strong>The Local Bridge:</strong> The extracted parameters are sent back to <strong>secure, on-premise backend</strong>.</li><li><strong>Data Sovereignty:</strong> Backend service retrieves the records and sends them directly to the frontend.</li></ul><p>The LLM is a glorified “Menu Navigator” — it knows the <em>intent</em>, but it never touches the <em>content</em>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kWjznlgamGmpEEWN3wKbRw.png" /><figcaption>Low-Risk Architecture</figcaption></figure><h3>Why This is High-ROI: Micro-Efficiencies Add Up</h3><p>High-ROI doesn’t have to mean “replacing a department.” It means making the department you have more faster and more accurate.</p><ol><li><strong>Zero Migration Cost:</strong> This feature sits <em>on top</em> of your existing infrastructure. It uses your current backend services and security layers. You aren’t rebuilding the house; you’re just adding a smart doorbell.</li><li><strong>Error Reduction:</strong> By utilizing the “Execute Same” logic, you eliminate the risk of a user mistyping a currency pair or an account number from a previous deal.</li><li><strong>Instant Adoption:</strong> There is no learning curve for a search bar. If a user can type a sentence, they can use the tool.</li></ol><p>This approach shows that AI is a good candidate for an <em>Intelligent Copy-Paste Assistant </em>in finance. By keeping the LLM focused on intent and the local system focused on data, you can deliver high-value automation while keeping the compliance team happy. I think the most valuable AI in finance is the one that almost never sees your data.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5f07a2ad776f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Closing the AI Adoption Gap with Low-Risk, High ROI Projects (Part 2): The Intelligent Office…]]></title>
            <link>https://rizahorasan.medium.com/closing-the-ai-adoption-gap-with-low-risk-high-roi-projects-part-2-the-intelligent-office-2a1e99dff7ac?source=rss-84be84f2acd4------2</link>
            <guid isPermaLink="false">https://medium.com/p/2a1e99dff7ac</guid>
            <category><![CDATA[enterprise-ai-adoption]]></category>
            <category><![CDATA[ai-workflow-automation]]></category>
            <category><![CDATA[ai-adoption]]></category>
            <category><![CDATA[fintech]]></category>
            <category><![CDATA[ai-in-banking]]></category>
            <dc:creator><![CDATA[rizahorasan]]></dc:creator>
            <pubDate>Tue, 21 Apr 2026 17:49:47 GMT</pubDate>
            <atom:updated>2026-04-21T17:49:47.041Z</atom:updated>
            <content:encoded><![CDATA[<h3>Closing the AI Adoption Gap with Low-Risk, High ROI Projects (Part 2): The Intelligent Office Navigator</h3><p>In my <a href="https://medium.com/@rizahorasan/closing-the-ai-adoption-gap-with-a-low-risk-high-roi-projects-in-finance-p1-51846ad2957f">previous post</a>, I explored how banks can use AI to turn messy, human-written text into clean data. Today, I want to look at a second high-impact, low risk use case: <strong>Internal Navigation.</strong> In the complex world of financial operations, “where is that file?” is a common question that costs hours every year.</p><p><strong>The Problem: The “Directory Labyrinth”</strong></p><p>Most financial enterprices operate across a massive network of shared drives, internal application links, and document gateways. Whether it is a “Daily Spot Report” or a “Swap Manual Upload” folder, these locations are often buried in obscure network paths like //fs-01/finance/reports/....</p><p>New employees struggle to find them, and experienced staff often keep “cheat sheets” that quickly become outdated. Traditional keyword search often fails because it requires the user to know the exact name of the folder or report.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/792/1*iUDdJpLFR4mslDRW81F7fA.png" /><figcaption>Where are the compliance reports?</figcaption></figure><p>I propose to deliver low-risk, high-impact “quick wins”, backed by <a href="https://medium.com/@rizahorasan/closing-the-ai-adoption-gap-with-intent-driven-architecture-959c06707efa">Intent-Driven Requirement Discovery</a> techniques. By introducing an <a href="https://medium.com/@rizahorasan/closing-the-ai-adoption-gap-with-intent-driven-architecture-959c06707efa">intent-driven command layer</a> on top of existing workflows, organizations can deliver low-risk, high-impact “quick wins”, while the deeper AI infrastructure is still being built.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*TzF5hmCAvFs8zgFbfWNvzw.png" /><figcaption>Intent analysis is always in place</figcaption></figure><p><strong>The Solution: A Semantic Search Assistant</strong></p><p>Our second prototype acts as an <em>Intelligent Navigator</em>. Instead of requiring a user to remember a specific path, they can simply ask the assistant in plain English.</p><ul><li><strong>The Knowledge Base:</strong> We maintain a simple, centralized “Resource Map” that lists every critical business function, its description, and its exact digital location.</li><li><strong>Intuitive Matching:</strong> When a user asks, <em>“Where do I drop the manual swap files?”</em>, the AI understands the intent.</li><li><strong>Instant Redirection:</strong> It cross-references the request against the map and instantly provides the precise network path: //manualuploadserver/swap.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1010/1*vcU_d6syU4oxdQsxUdD0pA.png" /><figcaption>Knowledge Map as a markdown file</figcaption></figure><h3>Why This is a “Low-Risk, High-ROI” Win</h3><h4><strong>1. Security First</strong></h4><p>This tool does not grant new permissions. It simply points users toward existing resources. Access to the actual network paths remains governed by the bank’s existing security protocols. If a user doesn’t have permission to open the folder the AI finds, the security wall stays up.</p><h4>2. Guardrailed Accuracy</h4><p>The system is designed to only provide information found within the official directory map. If a user asks for something that doesn’t exist in the registry, the assistant is trained to return a “No Match” response rather than guessing or making up a fake link.</p><h4>3. Operational Speed</h4><p>By turning a five-minute search into a five-second question, the ROI scales with every employee. It handles everything from:</p><ul><li><strong>Transaction Reports:</strong> Finding daily summaries or failed trade logs.</li><li><strong>Onboarding:</strong> Locating KYC document drop-zones.</li><li><strong>App Links:</strong> Finding deep links to specific screens like “Spot Entry” or “Swift Monitor”.</li></ul><p>I think , closing the AI adoption gap isn’t about building a “god-model” overnight; it’s about deploying high-ROI tools that solve real-world headaches like “where is that folder?”. By mapping natural language to a controlled list of related products and secure network paths, we eliminate the risk of the AI “guessing” wrong. This approach ensures that every interaction is grounded in your bank’s existing Resource &amp; Directory Map while respecting your established security permissions. It’s a low-risk strategy that shows AI can be an immediate ally for the back office — turning hours of digital archaeology into seconds of clarity and paving the way for more ambitious automation.</p><p>PoC implementation is on <a href="https://github.com/horasan/ai_quick_win_p1/blob/main/ai_wf/ai_short_answer/ai_short_answer.py">github, horasan</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2a1e99dff7ac" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Closing the AI Adoption Gap with a Low-Risk, High ROI Projects in Finance (p1)]]></title>
            <link>https://rizahorasan.medium.com/closing-the-ai-adoption-gap-with-a-low-risk-high-roi-projects-in-finance-p1-51846ad2957f?source=rss-84be84f2acd4------2</link>
            <guid isPermaLink="false">https://medium.com/p/51846ad2957f</guid>
            <category><![CDATA[ai-in-finance]]></category>
            <category><![CDATA[digital-transformation]]></category>
            <category><![CDATA[ai-adoption]]></category>
            <category><![CDATA[high-roi]]></category>
            <category><![CDATA[finance-app-development]]></category>
            <dc:creator><![CDATA[rizahorasan]]></dc:creator>
            <pubDate>Mon, 06 Apr 2026 20:43:34 GMT</pubDate>
            <atom:updated>2026-04-06T20:43:34.779Z</atom:updated>
            <content:encoded><![CDATA[<p>While organizations invest heavily in building compliant, AI-ready infrastructure, there is an AI adoption gap where users remain stuck in legacy workflows. This leaves powerful systems underutilized and delays real value realization.</p><p>I propose to deliver low-risk, high-impact “quick wins”, backed by <a href="https://medium.com/@rizahorasan/closing-the-ai-adoption-gap-with-intent-driven-architecture-959c06707efa">Intent-Driven Requirement Discovery</a> techniques. By introducing an <a href="https://medium.com/@rizahorasan/closing-the-ai-adoption-gap-with-intent-driven-architecture-959c06707efa">intent-driven command layer</a> on top of existing workflows, organizations can deliver low-risk, high-impact “quick wins”, while the deeper AI infrastructure is still being built.</p><p>The goal of this PoC application is to introduce an AI backed feature so that end users can continue to use the same flows, with fewer clicks and faster execution.</p><p><strong>Existing application</strong> has classic CRUD operations features. User opens a screen via a menu, enters the required information and clicks the Submit button.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KzbzpZsK_RIs9kxKYF0fTw.png" /><figcaption>Traditional Flow</figcaption></figure><p>AI backed application has an additional commandbar to enter the <strong>prompts</strong>. So that end users can enter a prompt like “Create a spot transaction for customer 15, EUR/USD, sell 10,000. Value date tomorrow”.</p><p>The only change in the application is the prompting textbox.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*aCzzeSVMUMWFnZGI21_lEQ.png" /><figcaption>AI Backed Flow</figcaption></figure><p>The system interprets this and <strong>pre-fills the transaction screen, </strong>nothing more, nothing less.</p><ul><li>No workflow changes</li><li>No system replacement</li><li>No disruption</li></ul><p>Just fewer clicks and faster execution.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tNsDQ4i4LyxIRVAfph-Feg.png" /><figcaption>Low risk — High ROI</figcaption></figure><h3><strong>Low Risk</strong></h3><p>This design deliberately avoids the biggest risks associated with AI in enterprise systems, <em>No Decision Responsibility for AI.</em></p><p>The AI does <strong>not</strong>:</p><ul><li>Execute transactions</li><li>Approve actions</li><li>Make business decisions</li></ul><p>It only <strong>suggests structured inputs</strong>.</p><p><strong>Human-in-the-loop</strong> pattern is here. The control remains with the user:</p><ul><li>The same UI is used</li><li>The same validation applies</li><li>The <strong>Submit button is still the final step</strong></li></ul><p>Nothing is automated beyond data entry.</p><p>No workflow disruption:</p><p>The existing process remains intact:</p><ul><li>Navigation</li><li>Validation</li><li>Approval chains</li></ul><p>All unchanged. The users still can choose to go with menu and manually entering the data.</p><p>Important Gap in this PoC:</p><ul><li><a href="https://rizahorasan.medium.com/audit-ready-ai-implementing-the-eu-ai-act-with-local-guardrails-and-langfuse-e7cd0e66120d"><strong>PII handling (masking, filtering, validation)</strong> </a>is not addressed in this PoC</li><li>This would be a critical requirement in production systems</li></ul><h3><strong>High ROI, Quick-Win</strong></h3><p>This is not a large transformation initiative. It’s a <strong>targeted enhancement</strong>.</p><p>No changes to Existing System</p><ul><li>No backend workflow redesign</li><li>No approval flow changes</li><li>No integration complexity</li></ul><h4>Immediate Productivity Gain</h4><ul><li>Removes repetitive data entry</li><li>Reduces navigation effort</li><li>Speeds up transaction preparation</li></ul><h4>Minimal User Training</h4><ul><li>Users stay in the same UI</li><li>No new tools to learn</li><li>Natural language is intuitive</li></ul><p>Adoption becomes almost effortless: <em>Nothing feels new, just faster.</em></p><h3><strong>Core Components</strong></h3><p>The system is intentionally simple and modular. Key characteristics are Structured outputs (predictable, typed data) and Controlled execution paths (no free-form AI behavior).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MRPHODZs7HeLooqnlQ8jbw.png" /><figcaption>Core Components</figcaption></figure><h4>Interface Layer</h4><ul><li>Captures user intent via natural language</li><li>Displays pre-filled transaction screens</li></ul><h4>API / Orchestration Layer</h4><ul><li>Handles communication between UI and backend</li><li>Keeps the system decoupled</li></ul><h4>Intelligence Layer (Agent Workflow)</h4><p>A controlled, state-machine-like pipeline:</p><ol><li><strong>Intent Classification</strong><br> Identify the transaction type</li><li><strong>Entity Extraction</strong><br> Extract structured data using schema constraints</li><li><strong>UI Mapping</strong><br> Map extracted data to UI fields</li></ol><h4><a href="https://medium.com/@rizahorasan/closing-the-ai-adoption-gap-with-intent-driven-architecture-959c06707efa"><strong>Intent Logging Layer</strong></a></h4><p>This layer acts as a feedback system that captures every (successful or not) user prompt and turns it into structured, analyzable data. Instead of discarding failed or partial interactions, the system records them as signals of unmet needs, emerging workflows, or UX friction.</p><h3><strong>Adding a New Product</strong></h3><p>This system is designed with the <strong>Open–Closed Principle</strong> in mind.</p><p>To add a new product:</p><ol><li>Define a new schema (what data is needed)</li><li>Add UI field mappings</li></ol><p>That’s it.</p><p>No need to:</p><ul><li>Change agent logic</li><li>Modify workflows</li><li>Rewrite APIs</li></ul><p>The system dynamically adapts based on configuration.</p><h3>Final Thought</h3><p>This approach demonstrates a pragmatic path to AI adoption:</p><ul><li>Start small</li><li>Focus on usability</li><li>Keep humans in control</li><li>Avoid disrupting what already works</li></ul><p>The result is not a radical transformation — but a <strong>high-impact improvement</strong> that delivers value immediately, with minimal risk. The PoC code is avaliable on <a href="https://github.com/horasan/ai_quick_win_p1">horasan@github</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=51846ad2957f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Closing the AI Adoption Gap with Intent-Driven Architecture]]></title>
            <link>https://rizahorasan.medium.com/closing-the-ai-adoption-gap-with-intent-driven-architecture-959c06707efa?source=rss-84be84f2acd4------2</link>
            <guid isPermaLink="false">https://medium.com/p/959c06707efa</guid>
            <category><![CDATA[ai-product-strategy]]></category>
            <category><![CDATA[ai-in-banking]]></category>
            <category><![CDATA[digital-transformation]]></category>
            <category><![CDATA[ai-adoption]]></category>
            <category><![CDATA[intent-engineering]]></category>
            <dc:creator><![CDATA[rizahorasan]]></dc:creator>
            <pubDate>Sun, 05 Apr 2026 21:24:28 GMT</pubDate>
            <atom:updated>2026-04-05T21:24:28.158Z</atom:updated>
            <content:encoded><![CDATA[<p>The problem I am proposing a solution is that in complex enterprise systems, end users cannot leverage AI capabilities because they don’t know what to ask or what is even possible. While organizations invest heavily in building compliant, AI-ready infrastructure, there is a gap where users remain stuck in legacy workflows. This leaves powerful systems underutilized and delays real value realization.</p><p>I propose to deliver low-risk, high-impact “quick wins”, backed by Intent-Driven Requirement Discovery techniques. By introducing an intent-driven command layer on top of existing workflows, organizations can deliver low-risk, high-impact “quick wins”, while the deeper AI infrastructure is still being built.</p><p><strong>Quick Wins and Passive Requirements Engineering</strong></p><p>Assume an existing UI with CRUD features, in finance, Spot Entry operation. Traditional workflow is simple: Navigate to FX module, select Spot Entry screen, search for customer, enter currency pair and enter amount.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KzbzpZsK_RIs9kxKYF0fTw.png" /><figcaption>Traditional CRUD Operaton</figcaption></figure><p>A quick-win, low risk AI feature might be giving the end user a textbox for prompts. End users can enter “Open spot entry for customer 15 for EUR/USD, side is sell and amount 1000 euro”. Then the related UI can be opened with the captured intent and key/value pairs.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*aCzzeSVMUMWFnZGI21_lEQ.png" /><figcaption>AI featured CRUD Operation</figcaption></figure><p><strong>The Hidden Advantage: Intent-Driven Requirement Discovery</strong></p><p>Intent-Driven Requirement Discovery is the systematic capture and analysis of user intent signals from natural language interactions, without explicit requirement gathering processes.</p><p>The most strategic aspect of this architecture is not execution. Observation is the real gain. Every user prompt becomes a <strong>data point of intent</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*jYQ900qDq0gbpGmfYfDB_Q.png" /><figcaption>Core Components &amp; Intent Logging Later</figcaption></figure><p><strong>Intent Logging Layer</strong> is the focused component. This layer acts as a feedback system that captures every (successful or not) user prompt and turns it into structured, analyzable data. Instead of discarding failed or partial interactions, the system records them as signals of unmet needs, emerging workflows, or UX friction. For example, if users repeatedly type “create recurring FX hedge for customer 42,” but no such feature exists, this indicates a clear product opportunity. Similarly, prompts like “sell euro dollar for client ABC” that fail due to ambiguous phrasing highlight gaps in intent recognition or terminology mapping. Even incomplete inputs such as “spot trade for 15” can reveal where users expect smart defaults or guided clarification. Over time, aggregating and classifying these prompts enables teams to identify high-frequency unmet intents, prioritize roadmap items, and continuously align the system with real user behavior. This approach can effectively transform raw interaction data into a living requirements engine.</p><p><strong>The Outcome</strong></p><p>When fully implemented, this system becomes more than a feature. It evolves into:</p><ul><li>A <strong>universal interaction layer</strong> across the application</li><li>A <strong>behavioral analytics engine</strong></li><li>A <strong>continuous product discovery mechanism</strong></li></ul><p>Instead of relying on workshops, surveys and stakeholder assumptions , organizations gain direct visibility on what users actually want to do and currently cannot.</p><p>I think Intent-Driven Requirement Discovery<strong> </strong>highlights a fundamental mismatch between system capability and user behavior. Intent-driven architecture resolves this by shifting interaction from navigation to expression. However, the true value lies beyond execution.</p><p>By capturing and structuring user intent, organizations can transform everyday interactions into a strategic asset. This approach might lead the continuous product evolution. This is not just a UX improvement. It is a new model for building software around real user needs.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=959c06707efa" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Experienced Colleague Who Never Sleeps: Scaling Senior Expertise via Microsoft Teams AI Agents]]></title>
            <link>https://rizahorasan.medium.com/the-experienced-colleague-who-never-sleeps-scaling-senior-expertise-via-microsoft-teams-ai-agents-f406852caf77?source=rss-84be84f2acd4------2</link>
            <guid isPermaLink="false">https://medium.com/p/f406852caf77</guid>
            <category><![CDATA[engineering-mangement]]></category>
            <category><![CDATA[microsoft-teams]]></category>
            <category><![CDATA[microsoft-365]]></category>
            <category><![CDATA[ai-productivity]]></category>
            <category><![CDATA[ai-agent]]></category>
            <dc:creator><![CDATA[rizahorasan]]></dc:creator>
            <pubDate>Sun, 01 Feb 2026 00:52:13 GMT</pubDate>
            <atom:updated>2026-02-01T00:52:13.780Z</atom:updated>
            <content:encoded><![CDATA[<p>In any high-performing team, experienced colleagues are the most expensive investment. This is not just in terms of compensation, but in the sheer opportunity cost of their time. When a seasoned expert is forced to pause deep, strategic work to answer a repetitive question about a standard process or historical context, the organization pays a “double tax”: the high hourly rate of the expert and the context-switching tax.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/696/1*qvesUyg03mJWth-pUJNg0Q.png" /></figure><p><strong>Expertise Gap: Lose-Lose Situation</strong></p><p>Crucial institutional knowledge is currently siloed within the minds of <strong>experienced colleagues</strong>, making them the unintended gatekeepers of progress. When a newer team member hits a roadblock, they require specific context to move forward, but the expert is often buried in high-stakes tasks with zero bandwidth for interruptions. This creates a cycle of inefficiency: the inexperienced member remains in a state of “blocked waiting,” while the seasoned expert eventually sacrifices their focus to handle repetitive, low-leverage queries. Even when documentation exists, static wikis are often too dense or disconnected to provide a direct answer to a specific problem. The result is a lose-lose situation where talent is stalled and expertise is drained by manual repetition.</p><p><strong>MS Teams as the experienced colleague</strong></p><p>The strategic value of using Microsoft Teams as the primary interface is rooted in its role as the modern “operating system” for the workplace. Because the vast majority of professional communication already flows through Teams chats, the AI agent isn’t a new destination. AI agents are a native participant in the existing conversations. For an experienced colleague, this means the agent acts as a first-line defender, intercepting queries directly within the same chat windows where interruptions usually occur. This proximity is vital; it eliminates “portal fatigue” and ensures that getting an answer doesn’t require leaving the flow of work. By turning a communication tool into a knowledge-delivery engine, the agent ensures that the gap between a question and a solution is closed instantly, right where the team is already collaborating.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/695/1*0Zf0VIjaeBVLZ1nOhr-CSw.png" /></figure><h3><strong>Microsoft 365 Agents Toolkit</strong></h3><p>Using the <strong>Microsoft 365 Agents Toolkit</strong>, we can move from “Tribal Knowledge” to “Agentic Knowledge”. Below you can find an implemantion for a “senior ” AI Teams Bot.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Y1UJYMxtwsxQdZr4R8kpgQ.png" /><figcaption>Technical AI Bot</figcaption></figure><p>This implementation is not only for technical workflows. The strategic value of an AI agent extends far beyond the engineering department;<br>it is a force multiplier for business users as well. In any organization, non-technical teams frequently run into “context walls” where they need specific data or process clarity to close a deal or finish a report. By serving as an accessible bridge to complex systems, the agent allows business users to query sophisticated domains using natural language. Instead of filing a ticket or waiting days for a specialized report, a business user can simply ask the agent in Teams for a summary of a relationship or a status update. <br>This democratization of information ensures that everyone, regardless of their technical background, can leverage the organization’s collective intelligence to make faster, data-driven decisions.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MAczp6hNLPnTaMyHd8lRAQ.png" /><figcaption>AI Agent Business user</figcaption></figure><p>Ultimately, deploying an AI agent within Microsoft Teams is more than just a technical upgrade; it is a fundamental shift in how an organization values its most precious resource , time. By capturing the unique insights of your <strong>experienced colleagues</strong> and making them instantly accessible to everyone, you transform “tribal knowledge” from a hidden liability into a scalable asset. You effectively stop paying your top talent to repeat themselves and start empowering your newer team members to be self-sufficient from day one. In the modern enterprise, the competitive edge belongs to the leaders who remove the friction of “waiting” and replace it with the speed of “knowing.” The tools are already in your chat window; it’s time to put them to work.</p><p>You can get the source code via <a href="https://github.com/horasan/ai_agent_in_ms_teams">horasan, github</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f406852caf77" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Audit-Ready AI: Implementing the EU AI Act with Local Guardrails and Langfuse]]></title>
            <link>https://rizahorasan.medium.com/audit-ready-ai-implementing-the-eu-ai-act-with-local-guardrails-and-langfuse-e7cd0e66120d?source=rss-84be84f2acd4------2</link>
            <guid isPermaLink="false">https://medium.com/p/e7cd0e66120d</guid>
            <category><![CDATA[compliance]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[fintech]]></category>
            <dc:creator><![CDATA[rizahorasan]]></dc:creator>
            <pubDate>Sat, 24 Jan 2026 23:48:08 GMT</pubDate>
            <atom:updated>2026-01-24T23:48:08.483Z</atom:updated>
            <content:encoded><![CDATA[<p>In 2026, the conversation around Generative AI in Europe has shifted. <br>Asking if an LLM is “powerful” is no longer enough to solve a business problem; companies are asking if it is “compliant” enough to be <em>deployed</em>. For critical sectors such as Banking, Healthcare, and Insurance, a 100s-billion parameter model is a liability if it lacks a verifiable audit trail. Under the EU AI Act, “Safety” is not just a system prompt — it is a technical requirement. <br>To bridge the gap between innovation and regulation, we must treat guardrails as first-class citizens in our architecture, moving beyond cloud-based safety APIs to a sovereign, local, and audit-ready designs.</p><p><strong>Requirement</strong>: Our customer-facing AI chatbot must interact with users <br>- without exposing PII to external vendors, <br>- providing hallucination-free answers, <br>- keeping banking context<br>- always safe taxonomy<br>- maintaining a full audit trail, <br>so that the bank remains fully compliant with the EU AI Act, GDPR and ready for AI Audit.</p><p><strong>A solution proposal</strong>: These requirements hit Art. 15: Robustness, Art. 13: Transparency, Art. 10: Data Governance, Art. 9: Risk Management, Art. 15: Accuracy, Art. 12: Record-keeping in the EU AI Act.</p><pre>| Guardrail Layer    | Local Model / Tool | Role in the Pipeline                                         | EU AI Act Alignment        |<br>|--------------------|--------------------|--------------------------------------------------------------|----------------------------|<br>| Input Integrity    | DeBERTa-v3         | Detects Prompt Injection &amp; Malicious Intent                  | Art. 15: Robustness        |<br>| Contextual Intent  | BART-MNLI          | Ensures query is within Banking scope                        | Art. 13: Transparency      |<br>| Data Privacy       | MS Presidio        | Automates PII detection &amp; pseudonymization                   | Art. 10: Data Governance   |<br>| Safety Taxonomy    | Llama-Guard 3      | Checks output against 14 risk categories                     | Art.  9: Risk Management   |<br>| Groundedness       | DeBERTa NLI        | Fact-checks response against internal data                   | Art. 15: Accuracy          |<br>| Traceability       | Langfuse (Local)   | Logs every trace and score for auditors                      | Art. 12: Record-keeping    |</pre><p>Below, the high level design is given.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/835/1*QtpE3AjGzzHcMmyt84JN7Q.png" /><figcaption>Langgraph Workflow with guardrails</figcaption></figure><p>Every node in the workflow is a specialized checkpoint, enabling individual logging and transparent examination of the entire AI decision-making process. By decoupling the safety layers into discrete nodes, we ensure full observability. Each component can be monitored and analyzed in isolation, simplifying both troubleshooting and compliance reporting.</p><ol><li><strong>Input Integrity: DeBERTa-v3</strong><br>The Challenge: Prompt injection and malicious intent. The Solution: Using DeBERTa-v3 fine-tuned for sequence classification allows us to intercept “jailbreak” attempts before they ever reach the core LLM. Unlike simple keyword filters, DeBERTa understands the semantic structure of a prompt, catching subtle “ignore previous instructions” attacks. EU AI Act Alignment: Art. 15 (Robustness &amp; Cybersecurity). High-risk systems must be resilient against attempts by unauthorized third parties to alter their use, outputs, or performance by exploiting system vulnerabilities.</li><li><strong>Contextual Intent: BART-MNLI</strong><br>The Challenge: Staying within the “Intended Purpose.” The Solution: We implement BART-MNLI for zero-shot classification. This ensures that if a user asks a banking bot for medical advice or a recipe, the system identifies it as “Out of Scope” and politely declines. EU AI Act Alignment: Art. 13 (Transparency &amp; Purpose Limitation). AI systems must be designed to ensure their operation is sufficiently transparent to enable users to interpret the system’s output and use it appropriately according to its intended purpose.</li><li><strong>Data Privacy: MS Presidio</strong><br>The Challenge: PII leaks in unstructured prompts. The Solution: Microsoft Presidio acts as our “Privacy Vault.” It scans incoming text for IBANs, names, and contact details, replacing them with generic placeholders (e.g., &lt;PERSON_1&gt;) before sending the data to the LLM. We reconstruct the personal data only on the “safe side” of the output. EU AI Act Alignment: Art. 10 (Data Governance). High-risk systems must be developed on the basis of training, validation, and testing data sets that meet quality criteria, including appropriate data governance and management practices.</li><li><strong>Safety Taxonomy: Llama-Guard 3</strong><br>The Challenge: Preventing toxic or harmful generated content. The Solution: We deploy Llama-Guard 3 (1b) as a specialized safety classifier. It evaluates the interaction against 14 risk categories, including financial crime, hate speech, and specialized advice. If the LLM generates a response that violates our policy, Llama-Guard blocks the output before it reaches the customer. EU AI Act Alignment: Art. 9 (Risk Management System). Providers must establish a continuous iterative process to identify, estimate, and mitigate known and foreseeable risks to health, safety, or fundamental rights.</li><li><strong>Groundedness: DeBERTa NLI</strong><br>The Challenge: Hallucinations and factual accuracy. The Solution: To prevent the AI from “inventing” banking policies, we use DeBERTa NLI as a cross-encoder. It compares the AI’s response against a verified internal knowledge base. If the response isn’t “grounded” in the provided facts, it is flagged as a hallucination. EU AI Act Alignment: Art. 15 (Accuracy &amp; Factuality). High-risk systems shall be designed and developed to achieve an appropriate level of accuracy, performing consistently throughout their lifecycle.</li><li><strong>Traceability: Langfuse (Local)</strong><br>The Challenge: Proving compliance during a regulatory audit. The Solution: Every decision made by the guards above is logged in a self-hosted instance of Langfuse. This “Flight Recorder” stores the original prompt, the masked prompt, the guardrail scores, and the final output. EU AI Act Alignment: Art. 12 (Technical Documentation &amp; Record-keeping). Systems must automatically record events (“logs”) over their lifetime to ensure a level of traceability appropriate to the intended purpose of the system.</li></ol><p>The era of “send the prompt and hope for the best” is over. By building a local, open-source safety stack (centered around specialized models like DeBERTa and Llama-Guard and unified by Langfuse ), we achieve more than just security. We achieve Data Sovereignty. We fulfill the strict requirements of the EU AI Act without paying a “cloud tax” or risking customer data egress. For the modern AI Products, these guardrails aren’t a hurdle; they are the foundation of a trust-first product.</p><p>The implementation can be found in <a href="https://github.com/horasan/ai_first_level_defence">github.com/horasan</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e7cd0e66120d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Beyond the Whiteboard: Turning Ideas into Verified Working Demos Before the Meeting Ends]]></title>
            <link>https://rizahorasan.medium.com/beyond-the-whiteboard-turning-ideas-into-verified-working-demos-before-the-meeting-ends-3d3d096dd3f1?source=rss-84be84f2acd4------2</link>
            <guid isPermaLink="false">https://medium.com/p/3d3d096dd3f1</guid>
            <category><![CDATA[llm-agent]]></category>
            <category><![CDATA[code-generation]]></category>
            <category><![CDATA[sdlc]]></category>
            <category><![CDATA[generative-ai-use-cases]]></category>
            <category><![CDATA[agentic-workflow]]></category>
            <dc:creator><![CDATA[rizahorasan]]></dc:creator>
            <pubDate>Mon, 19 Jan 2026 22:31:40 GMT</pubDate>
            <atom:updated>2026-01-19T22:31:40.182Z</atom:updated>
            <content:encoded><![CDATA[<h3>1. The High Cost of the “Follow-Up”</h3><p>In traditional software management, the space between a “great meeting” and a “working demo” is often a graveyard of productivity where momentum is sacrificed to the backlog. This “Follow-Up Gap” imposes a heavy translation tax, as every hour of delay between a whiteboard session <br>and the first commit increases the risk of misinterpreted requirements. <br>By failing to bridge the gap between ideas and execution in real-time, organizations pay a premium for a slow, error-prone process that often delivers the wrong solution.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*GEwZaDOBnT8LIa5ki475oQ.jpeg" /><figcaption>Generated by Nano Banana (via Gemini)</figcaption></figure><h3>2. Zero-Latency Development: The End of the Two-Week Prototype</h3><p>Zero-Latency Development transforms the project meeting from a precursor into the actual engine of the development cycle. By streaming raw notes into an agentic pipeline, the traditional two-week prototyping sprint is compressed into a sub-hour workflow that finishes before the attendees even leave the room. This shift replaces the “Guesswork Phase” of slide decks and mockups with an autonomous system that interprets strategic intent and initializes functional, real-time repositories. For management, this means moving instantly from high-level vision to a tangible, interactive demo, effectively eliminating the lead time between an idea and its first working iteration.</p><h3>3. The Business Case</h3><p>Imagine a meeting session where there isn’t a single developer in the room, only business stakeholders, product owners, and visionary leaders. As the group defines a new feature or logic flow, the meeting notes are uploaded to the AI Agent in real-time. By the time the “Next Steps” are being summarized, the agent has already interpreted the business logic, written the source code, and verified the functionality. Before the meeting officially ends, the stakeholders aren’t just looking at a summary of their ideas; they are interacting with a working MVP that was built while they were still talking.</p><h3>4. From “Next Week” to “Right Now”</h3><p>This Note-to-Code agent functions as an autonomous bridge between high-level requirements and running-MVP through a <strong>“no-touch” engineering pipeline</strong>. Unlike simple code assistants, this application utilizes a AI workflow to ingest unstructured notes and execute a full development lifecycle: orchestrating file architecture, generating type-hinted Python logic, and immediately triggering an automated <strong>verification loop via unit testing</strong>. By running these tests in a sandboxed subprocess, the agent identifies and self-corrects logic errors before the code ever reaches a human. The workflow culminates in a structured <strong>Pull Request (PR)</strong>, providing a seamless “human-in-the-loop” checkpoint where developers can review the AI’s verified output, ensuring that while the speed is autonomous, the final oversight remains firmly in human hands.</p><h3>5. Note-to-Code</h3><p>The AI workflow has 8 nodes, each for specific SDLC lifecycle.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/403/1*a1EL4EHGnLCpw5d-UwKe5A.png" /><figcaption>AI WF Nodes</figcaption></figure><p>When the agent starts, user provides <em>the meeting notes (plan.md)</em>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YB5j1OBkzpIjv9rw5QuvFg.png" /><figcaption>How to use Note-to-Code?</figcaption></figure><p>After the feature file from the plan.md is generated, related unittests are also generated. When the unittests are completed successfully, the implementation is sent to repo.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1guhTsAUDF_yrUasUtVr1w.png" /><figcaption>PR is generated automatically.</figcaption></figure><p>In this PoC, I have used “gemini-2.5-flash-lite”, which is not one of the most advanced models. But when the meeting notes are written in marked-down format, even this model gives very good results.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Lv0rJ3-rrJImmNlct6-iig.png" /><figcaption>gemini-2.5-flash-lite as code generator.</figcaption></figure><h3>6. The New ROI: From “Next Week” to “Right Now”</h3><p>The ultimate return on investment for Agentic Engineering isn’t measured in lines of code, but in the <strong>compression of the decision-making loop.</strong> When a product specification session turns directly into a live-code walkthrough, the entire physics of the business changes. The traditional “two-week wait” for a prototype is more than just a delay; it is a period of uncertainty where stakeholders lose focus and market windows close. By delivering a verified, working demo before the meeting adjourns, you eliminate the ambiguity of the “Follow-Up,” reduce the cost of experimentation to near-zero, and ensure that your engineering team is building exactly what the business needs. I think, the competitive advantage belongs to the leadership teams that stop taking action items and start delivering functional reality in real-time.</p><p>The implementation can be reached via <a href="https://github.com/horasan/ai_note_to_code">horasan, github</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3d3d096dd3f1" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[From Cost to Complience: Using Local LLMs as an Intellectual Guardrail for Privacy-by-Design in…]]></title>
            <link>https://rizahorasan.medium.com/from-cost-to-complience-using-local-llms-as-an-intellectual-guardrail-for-privacy-by-design-in-2494211c38fe?source=rss-84be84f2acd4------2</link>
            <guid isPermaLink="false">https://medium.com/p/2494211c38fe</guid>
            <category><![CDATA[eu-ai-act]]></category>
            <category><![CDATA[llm-agent]]></category>
            <category><![CDATA[generative-ai-use-cases]]></category>
            <category><![CDATA[langgraph]]></category>
            <category><![CDATA[ai-governance]]></category>
            <dc:creator><![CDATA[rizahorasan]]></dc:creator>
            <pubDate>Fri, 09 Jan 2026 21:18:39 GMT</pubDate>
            <atom:updated>2026-01-09T21:18:39.607Z</atom:updated>
            <content:encoded><![CDATA[<h3>From Cost to Complience: Using Local LLMs as an Intellectual Guardrail for Privacy-by-Design in Banking</h3><p>Most AI router nodes treat an ‘Others’ classification as a dead end. <br>In the high-stakes world of banking, I’ve come to see that intent isn’t a single data point, it’s more a negotiation. By moving the routing logic to a local LLM, I’ve created an intellectual guardrail that protects both the API budget and the customers’ privacy.</p><p><strong>A business case</strong>: A bank sets out to build an AI assistant with a clear mandate: help customers explore products, review transactions, manage personal information, and stay on top of notifications. For the other domains, assistant does not have any ability to help the customer.</p><p>The logic seems straightforward. If a customer asks about one of these areas, the system responds with the appropriate action. If not, the assistant politely asks a follow-up question to guide the conversation back on track. On paper, the scope is clean, bounded, and operational.</p><p><strong>A technical implementation</strong>: Under the hood, the solution takes shape as an agentic workflow. Every conversation begins at a routing node whose sole responsibility is to interpret user intent. Clear matches, like products, transactions, account details, notifications are passed downstream to specialized components. Everything else is treated as an exception, the others.</p><p><strong>The solution</strong>: “Everything else is treated as an exception”, but this is not an easy <em>else </em>statement. The router node must decide whether the user is still circling the banking domain with an imprecise request, or whether they have drifted entirely into unrelated territory. That single decision determines whether the system clarifies, redirects, or disengages. And it is where most real-world complexity quietly accumulates.</p><p>Multiple well-established theories explain this pattern. <strong>Social Presence Theory</strong>, often described as the “Sandbox Effect,” predicts that users will test an AI’s social and functional limits before trusting it with meaningful tasks. <strong>Information Foraging Theory</strong> explains why users issue vague or off-topic queries when the system’s capabilities are unclear: they are sampling the environment to discover what actions are possible. Finally, the <strong>Out-of-Distribution</strong> problem provides a hard baseline: industry data consistently shows that a significant share of production traffic does not align with narrow business logic, especially early in a session. Together, these frameworks show that <strong>the others</strong> group is not an anomaly but an expected outcome of human sense-making behavior interacting with constrained intent systems.</p><p>In practice, these theories converge on a single implication for system design: <em>a high volume of early </em><strong><em>others</em></strong><em> classifications is not noise to be eliminated, but signal to be interpreted</em>. Users are not failing to express intent; they are negotiating the interface, calibrating expectations, and testing whether the system can meaningfully respond. Treating these moments as errors leads to brittle routing and premature deflection. Treating them as part of the interaction lifecycle enables the agent to guide users from exploration to execution. In that light, <strong>others group</strong> is less a dead end and more a transitional state. This state reveals how humans actually approach AI systems before they commit to task-oriented language.</p><p>As a result, early-stage conversations are expected to generate low immediate business value, yet they still require engagement: the system must ask clarifying questions, surface its capabilities, and guide users toward actionable intent. This exploratory phase is inherently token-inefficient, producing little more than incremental context while the user formulates their request. Treating this as a cost problem is therefore reasonable. By routing these <strong>others</strong> or pre-intent interactions to a <strong><em>low-cost local model</em></strong>, the system absorbs exploratory traffic without incurring premium inference costs. Higher-capability models are reserved for moments when intent is explicit and downstream actions are justified. In this architecture, <strong>the local LLM</strong> is not a substitute for better design; it is a deliberate economic control applied to a predictable, low-signal phase of user interaction.</p><p>Beyond economic efficiency, this local routing architecture serves as a primary vehicle for compliance with the <strong>EU AI Act</strong>, specifically addressing the mandates for <strong>Article 15 (Accuracy, Robustness, and Cybersecurity)</strong>. In the financial sector, these kind of AI applications are often classified as <strong>high-risk</strong> where the Act demands that systems be “resilient as regards to errors, faults or inconsistencies.” By utilizing a local router to absorb and interpret the “others” group, the implementation enforces a <strong>deterministic gatekeeping layer</strong> that prevents out-of-distribution noise from reaching sensitive downstream logic. This “Fail-Safe” design fulfills the <strong>Article 9 Risk Management</strong> requirement by identifying and mitigating foreseeable misuse (like prompt injections or irrelevant data) at the earliest possible stage. Furthermore, the local deployment directly supports <strong>Data Minimization (GDPR &amp; AI Act Article 10)</strong>; by classifying and filtering intents within the bank’s own infrastructure, the system ensures that only the minimum necessary, non-sensitive data is passed forward, thereby reducing the regulatory attack surface and ensuring high-fidelity traceability for future audits.</p><p>According to these findings, a proposal solution might be as following.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/882/1*yOLeAhndAMYvHRSOj3KRgQ.png" /><figcaption>Workflow nodes and responsibilities</figcaption></figure><p>This architecture represents a <strong>multi-layered agentic guardrail system</strong> designed to balance high-precision banking logic with robust risk management. By utilizing a <strong>local LLM</strong> at the NODE_ROUTER, I establish an immediate, sovereign layer of defense that classifies intent without exposing sensitive customer data to external cloud APIs, directly aligning with <strong>EU AI Act Article 10</strong> regarding data governance. The most critical innovation here is the treatment of the NODE_OTHERS branch not as a failure state, but as a <strong>Transitional State</strong>. By routing ambiguous queries to a NODE_IS_IN_BANKING_DOMAIN check, the system differentiates between &quot;out-of-distribution&quot; noise (which is safely ended) and &quot;imprecise banking intent.&quot; This allows for a <strong>Clarification Loop</strong> that guides the user back to the router, ensuring that exploratory human behavior (&quot;Information Foraging&quot; or &quot;Sandboxing&quot;) is captured and redirected locally and cost-effectively. Finally, the use of a NODE_AGGREGATOR ensures that regardless of the path taken, the final output remains consistent, traceable, and ready for a compliance audit, fulfilling the <strong>Technical Robustness</strong> requirements of <strong>Article 15</strong>.</p><p><strong>Economic Guardrailing:</strong> Using a local 12B model (like Mistral NeMo) to “absorb” the high volume of exploratory or others traffic prevents wasting premium cloud tokens on low-value, ambiguous interactions.</p><p><strong>Privacy-by-Design:</strong> The local router acts as a PII (Personally Identifiable Information) firewall. Because the classification happens on-premises, you ensure that only verified, sanitized intents proceed to more complex downstream processing.</p><p><strong>Resilience to Human Vagueness</strong>: Instead of a brittle “Else/Error” statement, your architecture recognizes that users often negotiate the interface with vague language. The related_with_banking path provides a sophisticated &quot;Intent Recovery&quot; mechanism that improves user experience without sacrificing security.</p><p><strong>Regulatory Compliance:</strong> This flow provides a clear <strong>Audit Trail</strong>. I can prove exactly how a user’s prompt was handled, why it was categorized as others, and the logic used to determine if it stayed within the banking domain.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2494211c38fe" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Compliance by Design: Navigating High-Risk AI under the EU AI Act]]></title>
            <link>https://rizahorasan.medium.com/compliance-by-design-navigating-high-risk-ai-under-the-eu-ai-act-21491425bad6?source=rss-84be84f2acd4------2</link>
            <guid isPermaLink="false">https://medium.com/p/21491425bad6</guid>
            <category><![CDATA[eu-ai-act]]></category>
            <category><![CDATA[ai-governance]]></category>
            <category><![CDATA[responsible-ai]]></category>
            <category><![CDATA[finbert]]></category>
            <category><![CDATA[fintech]]></category>
            <dc:creator><![CDATA[rizahorasan]]></dc:creator>
            <pubDate>Mon, 29 Dec 2025 00:02:42 GMT</pubDate>
            <atom:updated>2025-12-29T00:02:42.588Z</atom:updated>
            <content:encoded><![CDATA[<p>A strategic blueprint for scaling responsible financial agents using local specialists and intelligent fallbacks.</p><p>When we set out to build an AI agent for the European financial market, the primary challenge wasn’t just coding a chatbot; it was navigating the strict requirements of the <strong>EU AI Act</strong>. In a landscape where financial AI is classified as ‘High-Risk,’ a standard ‘black box’ model is no longer a viable business option. We needed a system that could satisfy the Act’s demands for <strong>transparency, robustness, and human-in-the-loop oversight</strong> without exploding our cloud budget or slowing down customer response times. This is the story of how I used a <strong>‘Smart Triage’ architecture</strong> — pairing a fast, specialized local model with a high-reasoning LLM fallback — to create a system that is not only high-performing but <em>compliant</em> <em>by design</em>.</p><p><strong>The business requirement</strong>: “<em>In alignment with safety protocols, any output classified as a financial recommendation is automatically escalated for human review to ensure regulatory compliance.</em>”</p><p><strong>The solution</strong>: “A ‘Triage’ system using a specialized local model (FinBERT) to handle most of the traffic, with a targeted fallback to an LLM for “grey area” cases.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/613/1*adMvNd80_eHhvdceo9uuNA.png" /><figcaption>Local ML Model vs Expert LLM</figcaption></figure><p>This ‘Triage System’ acts as the critical final checkpoint within the NODE_OUTPUT_GUARD. Before any response reaches the user, the system runs a hybrid check using both a specialized local ML model and a high-reasoning LLM to determine if the content constitutes regulated advice. If the content is flagged as <em>recommendation, </em>the workflow directs the user to a human contact (NODE_END_FOR_OUTPUT_GUARD); otherwise, it proceeds to format the response for delivery. This checkpoint ensures that AI agent remains a helpful informational tool while strictly adhering to the <strong>EU AI Act’s</strong> safety and oversight requirements.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*21HAqWBi9TOUi55n9AUaZQ.png" /><figcaption>AI Workflow Architecture</figcaption></figure><p><strong>Local ML Model*</strong>: To maximize both speed and reliability, I utilized a specialized <strong>local ML model</strong> based on the <strong>FinBERT</strong> architecture, which was fine-tuned specifically to distinguish between general financial information and regulated advice. Unlike broad, “one-size-fits-all” AI, this model is a narrow specialist trained on synthetic** financial datasets (mostly generated with public LLMs) to understand the nuance of industry-specific language. By running this model locally, I eliminate the latency and high token costs associated with the public LLMs, allowing for near-instant classification. Most importantly, by focusing training on <strong>Precision</strong>, I’ve ensured that when the model identifies a ‘clear recommendation’ on the far right, the <strong>probability of a false positive is minimized</strong>. This high-precision foundation allows the ‘Smart Triage’ system to handle clear-cut cases with high reliability, while the <strong>0.35 safety threshold</strong> proactively captures ambiguous ‘grey area’ instances for escalation to the LLM, effectively mitigating the inherent risks of model error.</p><p>By integrating financial advice detection directly within the <strong>NODE_OUTPUT_GUARD</strong>, this architecture simultaneously satisfies the core &quot;High-Risk&quot; requirements of the <strong>EU AI Act</strong>. Specifically, the transition from this guard node to the <strong>NODE_END_FOR_OUTPUT_GUARD</strong> addresses <strong>Article 14 (Human Oversight)</strong>; by automatically terminating the AI&#39;s response and redirecting to a human contact when advice is detected, the system ensures a mandatory &quot;Human-in-the-Loop&quot; safety net. Furthermore, the hybrid combination of a specialized local ML model and a high-reasoning LLM fulfills <strong>Article 15 (Accuracy &amp; Robustness)</strong>. By implementing a conservative <strong>0.35 safety threshold</strong>, the system remains &quot;cautiously sensitive,&quot; minimizing the risk of unauthorized financial guidance while maintaining the high performance and resilience required for high-stakes financial applications.</p><p>*Local ML model can be reached on <a href="https://github.com/horasan/ai_is_financial_advice">horasan, github</a>.</p><p>**While synthetic data is a powerful tool for overcoming privacy constraints and data scarcity, it carries the inherent risk of <strong>model collapse</strong>, where the AI begins to “forget” real-world nuances by training on its own simplified outputs. This can lead to <strong>amplified biases</strong> and the loss of critical edge cases, as artificial datasets often lack the messy, unpredictable variety found in genuine human behavior. To maintain the integrity of our financial advice detection, we must treat synthetic data as a starting point rather than a final source, ensuring rigorous validation against real-world scenarios to prevent “semantic drift” in our compliance logic.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=21491425bad6" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>