<?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 Anantdhavale on Medium]]></title>
        <description><![CDATA[Stories by Anantdhavale on Medium]]></description>
        <link>https://medium.com/@anantdhavale?source=rss-b845d0c5e68e------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*YulaM1sJaHsAmJet59HU9w.png</url>
            <title>Stories by Anantdhavale on Medium</title>
            <link>https://medium.com/@anantdhavale?source=rss-b845d0c5e68e------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 04:43:01 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@anantdhavale/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[Design-Time Truth vs. Runtime Reality: How to Detect Process Divergence]]></title>
            <link>https://medium.com/@anantdhavale/design-time-truth-vs-runtime-reality-how-to-detect-process-divergence-3027ffb1ab92?source=rss-b845d0c5e68e------2</link>
            <guid isPermaLink="false">https://medium.com/p/3027ffb1ab92</guid>
            <category><![CDATA[work]]></category>
            <category><![CDATA[model-context-protocol]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[process-improvement]]></category>
            <dc:creator><![CDATA[Anantdhavale]]></dc:creator>
            <pubDate>Wed, 15 Jul 2026 17:00:58 GMT</pubDate>
            <atom:updated>2026-07-15T17:05:53.132Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="A biker is biking, they are wearing a bright red wind-cheater jacket. They are wearing a helmet and riding the bike in half standing position, at the backdrop we see a blurred city scape and a road." src="https://cdn-images-1.medium.com/max/1024/1*vJRp8nKLpHdfX4JM-DIuzw.avif" /></figure><p>Process divergence is the gap between how a process was designed to run and how it actually runs in production. Every documented process is design-time truth: accurate the day it was approved, decaying from that day forward as prompts change, people adapt, systems get patched, and exceptions harden into habits. Divergence is not a failure of documentation. It is the default behavior of any live process, and it is detectable, continuously, once the design exists as a machine-checkable standard.</p><p>This article defines process divergence, explains why it stays invisible under normal governance, and describes how continuous divergence detection works.</p><p><strong>What is process divergence?</strong></p><p>Process divergence is the measured distance between a process’s documented design and its observed execution. The design says approvals above a threshold require a second sign-off. Runtime reality shows a fraction of them proceeding on one. The design says exceptions escalate within 48 hours. Runtime reality shows a growing tail quietly aging past that window. Neither is a system outage or a permission breach. Both are the process no longer being the process.</p><p>The objection this concept answers is a fair one: “processes are design-time truth, so what good is documenting them?” The answer is that a documented design is only useless if nothing compares reality against it. The moment the design becomes a standard that live behavior is checked against, its decay becomes a signal instead of a liability.</p><p><strong>Why does divergence stay invisible?</strong></p><p>Because nothing in a typical stack is responsible for noticing it.</p><p>Access controls check permissions, and divergence rarely violates one. Monitoring checks system health, and a divergent process runs on perfectly healthy systems. Periodic audit checks samples, and divergence that concentrates outside the sample, or accumulated after the last cycle, is invisible by construction. Meanwhile the people running the process are the source of much of the divergence, adapting to real conditions faster than documentation updates, so they experience it as “how we do it now,” not as a gap worth reporting.</p><p>Agents accelerate all of this. Every prompt edit is an undocumented process change. Every added tool widens what the process effectively does. Agent-driven divergence accumulates at edit speed, not at the speed of organizational habit.</p><p><strong>How is process divergence detected continuously?</strong></p><p>Continuous divergence detection requires three components: a structured standard, an observable event stream, and per-action comparison between them.</p><p><strong>A structured standard.</strong> The designed process, its sequence, conditions, thresholds, and boundaries, represented in machine-checkable form. Prose SOPs cannot serve here; the design must answer yes-or-no questions about specific actions.</p><p><strong>An observable stream.</strong> The process’s actual execution as events: actions taken, by which actor, in what order, with what parameters. For agent-executed work this stream exists natively. For human-executed work it comes from the systems the work passes through.</p><p><strong>Per-action comparison.</strong> Each observed action is scored against the standard as it happens: in sequence or out of it, inside thresholds or past them, on a defined path or off one. Individual mismatches become flags. Patterns of mismatch become divergence findings: this step is now routinely skipped, this threshold is effectively higher than designed, this exception path has become the main path.</p><p>The output is not an alert per anomaly but a live answer to a governance question audits can only answer retrospectively: where, today, does this process differ from its design?</p><p><strong>What do you do with a divergence finding?</strong></p><p>A divergence finding forces exactly one of two decisions, and both are healthy: fix the behavior, or fix the design.</p><p>Some divergence is defect: a control being skipped, a boundary being crossed, drift that must be corrected back to design. Some divergence is information: the field found a better path, the threshold was set wrong, the design was never realistic. Continuous detection does not presume which; it surfaces the gap with specifics and hands the judgment to the process owner. What it removes is the third, silent option that dominates today, where the gap is neither corrected nor adopted, just unknown.</p><p>Over time this loop changes what documentation is. The design stops being a compliance artifact written at rollout and becomes a living standard, updated deliberately when reality wins the argument, enforced deliberately when it shouldn’t.</p><p><strong>Why does this matter more with agents in the loop?</strong></p><p>Because agents both cause divergence faster and make it detectable for the first time at full coverage.</p><p>Human process divergence was always sampled: you learned about it in audits, incidents, or exit interviews. Agent-executed processes emit their complete action stream, so comparison against the design can cover every action, not a periodic sample. The same shift that makes divergence accumulate at edit speed also makes 100 percent detection feasible. Organizations that structure their process designs get both halves; organizations that don’t get only the acceleration.</p><p><strong>Summary</strong></p><p>Design-time truth decays; that is not a flaw of documentation but the nature of live processes, and it grows faster once agents execute and modify workflows. Process divergence, the measured gap between designed and observed behavior, stays invisible to permissions, monitoring, and sampled audits because none of them compares reality to design. Continuous divergence detection makes the comparison per action, against a structured standard, turning the decay into a signal with only two outcomes: correct the behavior or update the design. The silent third outcome, an unknown gap, is the one it eliminates.</p><p><strong>Frequently asked questions</strong></p><p><strong>Is divergence detection the same as process mining?</strong> They are related but differ in timing and reference. Mining reconstructs the actual process from logs, typically retrospectively. Divergence detection compares live actions against a declared design, continuously, and flags gaps as they form.</p><p><strong>Is all divergence bad?</strong> No. Divergence is evidence of a gap; the judgment of defect versus improvement belongs to the process owner. The failure mode is not divergence itself but divergence that nobody knows about.</p><p><strong>Does this work for human-executed processes or only agents?</strong> Both, wherever the work leaves an event trail in systems. Agent-executed processes simply provide the most complete stream.</p><p><strong>How does this relate to semantic drift?</strong> Semantic drift is divergence at the level of a single agent’s purpose. Process divergence is the same gap at the level of the workflow. Detecting both uses the same principle: a declared standard, compared against observed behavior, continuously.</p><p><em>Homer Semantics turns your documented processes into live, machine-checkable standards, so the gap between design and reality becomes a signal you see, not a surprise you audit. Write to </em><a href="/service/mailto:info@homersemantics.com"><em>info@homersemantics.com</em></a><em> for a working session on one of your processes.</em></p><p>Visit <a href="https://www.homersemantics.com/">https://www.homersemantics.com/</a> to know more.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3027ffb1ab92" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Rise of the Product Factory: What Comes After LLM Model Progress Normalizes]]></title>
            <link>https://medium.com/@anantdhavale/the-rise-of-the-product-factory-what-comes-after-llm-model-progress-normalizes-6c29c70cb0b1?source=rss-b845d0c5e68e------2</link>
            <guid isPermaLink="false">https://medium.com/p/6c29c70cb0b1</guid>
            <category><![CDATA[fintech-startups]]></category>
            <category><![CDATA[ai-start-up]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[money]]></category>
            <category><![CDATA[software-engineering]]></category>
            <dc:creator><![CDATA[Anantdhavale]]></dc:creator>
            <pubDate>Thu, 09 Jul 2026 09:59:02 GMT</pubDate>
            <atom:updated>2026-07-09T09:59:02.714Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="A factory with green colored thread-yarn is seen. It is used as a symbol of the software factory model described in the article." src="https://cdn-images-1.medium.com/max/612/1*KIn31Pc_AU4PvcscU4B52Q.jpeg" /></figure><p>I want to call it the Product Factory thesis. Nobody has named it yet, but this behavior is everywhere: since AI has commoditized software production, startups are launching products continuously against evolving customer needs instead of scaling one, treating the single-product SaaS company as obsolete. The thesis is half right. Parallel product launching is a real and growing behavior.</p><p>Parallel product scaling, however, remains rare, and the companies that appear to do it are usually running one shared asset behind many front-ends.</p><p>This article examines why the thesis is surging now, what the evidence actually supports, and what replaces the product as the unit of defensibility when building one stops being hard.</p><p><strong>Why is the Product Factory pattern surging now?</strong></p><p>Because the market has stopped pricing model progress as the story, and attention is moving down the stack to who captures value in the application layer.</p><p>Frontier LLM improvements, once the defining narrative of the AI trade, have normalized: each release is absorbed as expected rather than treated as an event. When the layer above stops being the differentiator, the question becomes what happens in the layer below, where models are inputs. There, one fact is undeniable: the cost of producing working software has collapsed. Teams ship in weeks what took quarters. From that fact, the Product Factory draws its conclusion: if products are cheap, make many of them.</p><p><strong>How is the pattern reflecting in the market?</strong></p><p>Market trends show parallel launching as a search strategy, followed by classic concentration once something works.</p><p>Startups today routinely run <strong>several products simultaneously</strong>. But this is best read as price discovery: launching three products is the new A/B test, feasible for the first time because a launch costs weeks, not quarters. Track the same companies twelve to eighteen months on and the pattern is consolidation behind the winner. The parallel phase is how they search; it is not how they scale. The observation “everyone is launching parallel products” is largely an artifact of watching a young cohort mid-search.</p><p>The longer record points the same direction. Venture studios and product factories have systematically underperformed focused companies, because each product carries its own distribution, support, and trust-building, and none of those commoditized. A factory of ten products is usually ten go-to-market problems sharing one exhausted team.</p><p><strong>Where does the factory model genuinely work?</strong></p><p>In exactly two configurations, and both are less “many products” than they appear.</p><p><strong>Shared distribution.</strong> A studio with one audience, marketplace, or channel that every new product plugs into. The products vary; the customer acquisition asset is singular. Each launch rides distribution that already exists.</p><p><strong>Shared engine.</strong> A company whose products are thin surfaces over one underlying capability: one data asset, one platform, one core technology, expressed as multiple offerings for multiple buyers.</p><p>In both cases the honest description is one asset, many front-ends. The Product Factory is not producing independent products; it is amortizing a moat. Factories without a shared asset, where each product must earn its own distribution and trust from zero, are the configuration the evidence keeps rejecting.</p><p>If building is commoditized, what is actually scarce?</p><p>Three things, and cheap production makes each of them more valuable, not less.</p><p><strong>Distribution.</strong> When anyone can build anything, the constraint moves entirely to reaching buyers. Owned channels, installed relationships, and category association appreciate as build costs fall, the same way cheap content made brands more valuable rather than less.</p><p><strong>Trust and depth.</strong> Enterprise buyers do not purchase software; they purchase confidence that it will work inside their constraints. That confidence is accumulated per product, per segment, slowly, and cloning a feature set does not clone it.</p><p><strong>Learning rate.</strong> The scarcest pre-traction resource is iterations against one problem. A team splitting attention across five products runs a fifth of the learning cycles on each. Commoditized building shortens each cycle; it does not multiply the team’s capacity to absorb what the cycles teach.</p><p><strong>So what actually comes after the LLM model race?</strong></p><p>The defensible sequence is wedge first, factory second: concentrate until one product owns a customer relationship, then use commoditized building to expand across that relationship fast.</p><p>This is land-and-expand with a radically cheaper expand. The first product does the expensive, uncommoditizable work of earning distribution and trust in a segment. Every subsequent product is then launched into an existing relationship instead of a cold market, which is precisely the configuration where factory economics turn positive. “Build many products because building is cheap” fails as a starting strategy and excels as a scaling strategy. The order is the entire argument.</p><p><strong>Summary</strong></p><p>The Product Factory is a name for a pattern already underway, and the pattern mistakes a fallen cost for a vanished bottleneck. Building was never the bottleneck; distribution, trust, and learning rate were, and cheap production concentrates their value further. What the evidence supports is narrower and more useful: parallel launching as cheap search, consolidation behind what works, and factory-style expansion only on top of a shared asset, an owned channel, an owned engine, or an owned customer relationship. As LLM progress normalizes and attention shifts to the application layer, the winners will not be the companies that ship the most products. They will be the companies that earned one wedge and then let cheap building compound it.</p><p>Frequently asked questions</p><p><strong>Is the single-product SaaS company really over?</strong> As a permanent end-state, increasingly yes: successful companies expand into adjacent products faster than before. As a starting point, no: the wedge product still has to win first, and splitting focus before it does remains the classic failure mode.</p><p><strong>Doesn’t fast cloning make focus riskier, since one product can be copied?</strong> Cloning copies features, not distribution or trust. A focused product that owns its segment’s relationship is harder to displace than a portfolio of shallow ones, precisely because the copyable part was never the moat.</p><p><strong>What about AI studios shipping dozens of small apps profitably?</strong> They typically monetize one shared asset, an audience, a channel, or an engine, through many surfaces. That model works, but it is one moat with many faces, not many independent products.</p><p><strong>How should an early-stage company apply this?</strong> Use cheap building for search: test in parallel, briefly and honestly. Then concentrate everything behind the product that finds pull, and defer the factory until that product has earned the distribution the rest of the portfolio will ride on.</p><p>Notes :</p><p><em>Check our consulting and implementation services at </em><a href="https://procors.com/"><em>https://procors.com/</em></a><em><br>Know how our Process Context Engine helps organization unlock the value from business knowledge assets at </em><a href="https://www.homersemantics.com/"><em>https://www.homersemantics.com/</em></a></p><p>This article was written with the help of the Claude model.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6c29c70cb0b1" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[93% Say Insights Need Business Context. Almost Nobody Has Built the Context Layer.]]></title>
            <link>https://medium.com/@anantdhavale/93-say-insights-need-business-context-almost-nobody-has-built-the-context-layer-12d364e5b479?source=rss-b845d0c5e68e------2</link>
            <guid isPermaLink="false">https://medium.com/p/12d364e5b479</guid>
            <category><![CDATA[money]]></category>
            <category><![CDATA[ai-context-management]]></category>
            <category><![CDATA[revenue]]></category>
            <category><![CDATA[context-api]]></category>
            <category><![CDATA[business-strategy]]></category>
            <dc:creator><![CDATA[Anantdhavale]]></dc:creator>
            <pubDate>Thu, 09 Jul 2026 09:52:41 GMT</pubDate>
            <atom:updated>2026-07-09T09:52:41.607Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*dnqjmegzvs-nwWOES9VuZA.jpeg" /></figure><p>The business context gap is the distance between having data and having data that means something inside your operations. <strong>Salesforce’s State of Data &amp; Analytics report (2nd edition),</strong> a survey of more than 7,600 business and data leaders worldwide, quantifies it: 93% of business leaders agree insights are only relevant when grounded in business context, while 49% of data and analytics leaders admit their companies occasionally or frequently draw incorrect conclusions from data that misses or misunderstands that context.</p><p>Nearly everyone agrees context is the prerequisite. Nearly half concede their organization gets it wrong anyway. This article examines why that gap persists and what actually closes it.</p><p><strong>What does the data say about the context gap?</strong></p><p>It says organizations have gotten more data-driven without getting more context-driven, and the two are diverging.</p><p>Per the Salesforce survey, 63% of business leaders now describe their organizations as very data-driven, up ten points from 2023. Yet almost the same share of technical leaders, 63%, say their companies struggle to drive business priorities with data, and 42% of business leaders say their data strategies don’t fully align with business objectives. Data leaders estimate roughly a quarter of their organizations’ data is untrustworthy outright.</p><p>The pattern is consistent: investment in data volume and access has outpaced investment in the layer that makes data interpretable, the business context that says what a number means, which process it belongs to, and what should follow from it.</p><p><strong>Why does missing context produce wrong conclusions?</strong></p><p>Because data without context forces every consumer, human or AI, to supply their own interpretation, and interpretations diverge.</p><p>A revenue figure means something different mid-quarter versus post-close. A flagged transaction means something different inside a remediation process versus outside one. When that surrounding knowledge lives in people’s heads instead of in a queryable layer, each analyst, each department, and now each AI agent reconstructs it independently. The survey shows the downstream cost: 36% of business leaders have provided a data point that differed from a colleague’s, 32% have omitted questionable data points, and 32% have fallen back on gut-based decisions because the data they had couldn’t be trusted to mean what it appeared to mean.</p><p>None of these are data quality failures in the pipeline sense. The number was often correct. What was missing was the context to interpret it consistently.</p><p><strong>Why do AI agents make the context gap urgent?</strong></p><p>Because agents consume data at machine speed and supply their own interpretation with total confidence, so a context gap that produced occasional human error now produces systematic machine error.</p><p>The survey shows AI has already become the most common way business leaders find, analyze, and interpret data, ahead of self-service and technical help. It also shows what happens when those AI layers sit on ungrounded data: 89% of data and analytics leaders using AI report having experienced inaccurate or misleading outputs, and 84% agree AI’s outputs are only as good as its data inputs. More than half of companies training or fine-tuning their own models say they have wasted significant resources doing so on bad data.</p><p>An analyst who receives a context-free number might hesitate. An agent doesn’t. It acts on its best reconstruction of what the number means, every time, at scale. The gap that was a friction cost becomes an error generator.</p><p><strong>Why hasn’t the context layer been built?</strong></p><p>Because context has always been treated as documentation and training, not as infrastructure.</p><p>Business context lives in SOPs, process diagrams, onboarding decks, and the accumulated judgment of experienced staff. All of it is written for people. None of it is queryable by systems. The Salesforce data shows where organizations put their effort instead: real-time monitoring is the only data quality tactic with majority adoption at 69%, while data lineage tracking sits at 38%, and only 43% of data and analytics leaders have established formal governance frameworks at all, even as 88% say AI demands new approaches.</p><p>Monitoring tells you the pipeline is flowing. It cannot tell you whether the data flowing through it will be interpreted correctly on arrival, because interpretation depends on the layer nobody has built: business context structured as something a system can check against, not prose a person may or may not have read.</p><p><strong>What does building the context layer actually involve?</strong></p><p>Three moves separate organizations that close the gap from those that keep widening it.</p><p><strong>Structure the operating knowledge, not just the data.</strong> The processes, definitions, thresholds, and exception rules that give data its meaning must be represented in machine-queryable form, so a system consuming a number can also retrieve what that number means here, now, in this workflow.</p><p><strong>Ground AI consumption in that layer.</strong> Every agent or AI-assisted analysis should be able to check its interpretation and its intended action against the structured context before acting, rather than reconstructing meaning from a prompt.</p><p><strong>Treat context as versioned infrastructure.</strong> Business context changes as processes change. A context layer that isn’t maintained becomes the same stale SOP problem in a new format. It needs ownership and change control like any production system.</p><p><em>Building the organizational context is not one single thing — but a network of services.</em></p><p><strong>Summary</strong></p><p>The Salesforce data describes a consensus and a failure in the same breath: 93% agree insights require business context, half concede their organizations draw wrong conclusions without it, and the arrival of AI agents converts that gap from a human friction cost into a machine-scale error source. Organizations responded to the data era by building pipelines, warehouses, and monitoring. The context era requires something none of those provide: the operating knowledge of the business, structured as a layer systems can query, so that what a number means stops depending on who, or what, happens to be reading it.</p><p><strong>Frequently asked questions</strong></p><p><strong>Isn’t business context what a semantic layer in a BI tool provides?</strong> Partially. BI semantic layers define metrics and dimensions for reporting. Business context in the broader sense also includes process knowledge: sequences, conditions, exceptions, and boundaries that govern what should happen, which reporting-oriented semantic layers don’t capture.</p><p><strong>Is this a data quality problem?</strong> It overlaps but isn’t identical. The survey distinguishes them: organizations can have accurate data and still draw wrong conclusions because the interpretation layer is missing. Quality fixes the number; context fixes what the number means.</p><p><strong>Where do the statistics in this article come from?</strong> All figures are from Salesforce’s State of Data &amp; Analytics report, 2nd edition, a double-anonymous survey of more than 7,600 data, analytics, and line-of-business leaders across 17 countries and 18 industries.</p><p><strong>Where should an organization start?</strong> With one workflow where misinterpretation is expensive. Structure its context, definitions, process rules, thresholds, ground the AI touching that workflow in it, and expand from there.</p><p>Building the context layer is not limited to one specific thing — but we help organizations get going with building the “missing” business context : your processes, definitions, and rules, structured into a standard that both people and AI agents can query</p><p>— — —</p><p>Notes:</p><p>Write to <a href="/service/mailto:info@homersemantics.com">info@homersemantics.com</a> to start with one workflow.</p><p><em>Know how our Process Context Engine helps organization unlock the value from business knowledge assets at </em><a href="https://www.homersemantics.com/"><em>https://www.homersemantics.com/</em></a></p><p>Check our consulting and implementation services at <a href="https://procors.com/">https://procors.com/</a></p><p><em>This post is based on </em><strong>Salesforce’s State of Data &amp; Analytics report (2nd edition) and is written with the help of the Claude agent.</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=12d364e5b479" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[LangChain-agent-governance:Runtime checks for LangChain agents]]></title>
            <link>https://medium.com/@anantdhavale/langchain-agent-governance-runtime-checks-for-langchain-agents-4cbb97106a55?source=rss-b845d0c5e68e------2</link>
            <guid isPermaLink="false">https://medium.com/p/4cbb97106a55</guid>
            <category><![CDATA[langchain-agents]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[langchain]]></category>
            <category><![CDATA[langchain-tools]]></category>
            <dc:creator><![CDATA[Anantdhavale]]></dc:creator>
            <pubDate>Sat, 16 May 2026 16:46:53 GMT</pubDate>
            <atom:updated>2026-05-16T16:53:51.115Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/837/1*pJohExsv2iKQJ0yUWJ6x6A.jpeg" /></figure><p>LangChain makes it easy to build agents that can hit APIs and run workflows. But most teams tend to rely on prompts and hope the model behaves.</p><p>This package basically adds a runtime check layer in front of the LangChain tool execution.</p><p>What does it do?</p><p>1. wrap LangChain tools before they run</p><p>2. register or reuse a Cerone agent identity</p><p>3. sends the tool name and parameters for validation</p><p>3. gets back a decision</p><p>. approved</p><p>. rejected</p><p>. flagged</p><p>4. Then either lets the tool run, raises a flagged-action error, or blocks execution.</p><p>Beyond tool wrapping, it supports :</p><p>agent-level toolset setup<br>agent presets like coding_agent, research_agent, support_agent<br>bound workflow metadata like session_id, workflow_id, workflow_step<br>child-agent spawning for multi-step workflows<br>delegated tokens for child workflow steps<br>configurable handling for flagged action</p><pre>from langchain_core.tools import tool<br>from cerone import CeroneClient<br>from cerone_langchain import govern_agent_tools<br><br>@tool<br>def read_file(path: str) -&gt; str:<br>    with open(path, &quot;r&quot;, encoding=&quot;utf-8&quot;) as f:<br>        return f.read()<br><br>client = CeroneClient(api_key=None)<br><br>governed_tools, governor = govern_agent_tools(<br>    tools=[read_file],<br>    client=client,<br>    preset=&quot;coding_agent&quot;,<br>    capabilities=[&quot;file_read&quot;],<br>    environment=&quot;development&quot;,<br>    workflow_id=&quot;wf_repo_review&quot;,<br>    workflow_step=&quot;inspect_repository&quot;,<br>)<br><br>print(governed_tools[0].invoke({&quot;path&quot;: &quot;README.md&quot;}))</pre><p>Repo:</p><p><a href="https://github.com/AnantDhavale/cerone-langchain">https://github.com/AnantDhavale/cerone-langchain</a></p><p>Package name:</p><p>langchain-agent-governance</p><p>Appreciate feedback from people using LangChain for:</p><ul><li>coding agents</li><li>research agents</li><li>multi-step workflows</li><li>multi-agent systems</li></ul><p>The key thing I am trying to learn:</p><p>is the right abstraction here “tool checks”, “agent checks”, or “workflow checks” for real production LangChain users?</p><p>Note — The plugin works on Cerone, the runtime agent governance layer. It’s on PyPI ( pip install cerone). Cerone is not tied to any framework and can be used anywhere, such as your n8n workflows.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4cbb97106a55" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Agent Governance]]></title>
            <link>https://medium.com/@anantdhavale/agent-governance-69dba0bded66?source=rss-b845d0c5e68e------2</link>
            <guid isPermaLink="false">https://medium.com/p/69dba0bded66</guid>
            <category><![CDATA[ai-agent-governance]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[governance]]></category>
            <category><![CDATA[ai-governance]]></category>
            <dc:creator><![CDATA[Anantdhavale]]></dc:creator>
            <pubDate>Wed, 22 Apr 2026 07:00:33 GMT</pubDate>
            <atom:updated>2026-04-23T09:39:32.907Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="A Robot sits in an office and works. This article is about AI giovernance." src="https://cdn-images-1.medium.com/max/960/1*ZsdX18YhjO1r5oVWeFkT3A.jpeg" /></figure><p>AI Governance has become a foundational requirement as AI systems move production-scale usage, and start performing complex tasks.</p><p>At a high level, AI usage in applications can be divided into two types:</p><ul><li><strong>Inference calls</strong> (single LLM request/response, embedded in a traditional app flow)</li><li><strong>Agentic systems</strong> (LLM drives multi-step reasoning, tool use, or orchestration with some autonomy)</li></ul><p>The governance needs for these two differ significantly when it comes to control, observability, and risk. Embedded AI requires guardrails around prompt design, output validation, and cost/runtime monitoring. AI agents require deeper oversight, including decision traceability, tool usage control, intent adherence and bounded autonomy.</p><p>However, it is important to understand that AI governance cannot remain static or design-time only, it must operate at runtime. Runtime governance ensures that every interaction, decision, and deviation is monitored and controlled in real time.</p><p>This is exactly why we built G<strong>Core </strong>and <strong>Cerone</strong>.</p><p><strong>GCore (Governance Core) </strong>provides runtime governance for LLM-based embedded AI systems.</p><p><a href="https://pypi.org/project/homer-sdk/">https://pypi.org/project/homer-sdk/</a></p><p>Core allows operational teams to define their own context and runtime policies. It wraps your existing LLM client in a single line; as a result API calls stay identical, but every request now passes through a semantic control layer and every response goes into an immutable ledger. It does not require any rewrites, or new infra. Core :</p><ul><li>Works with OpenAI, Anthropic, Azure, or any local model</li><li>Drop-in:homer.openai()replacesopenai.OpenAI()</li><li>Policies in Python: no config files, no YAML</li><li>Enriched response: classification, norms, confidence, violations</li></ul><p>Consider a scenario — you have 30 teams calling LLMs in 30 different ways, you need consistent controls without becoming a bottleneck. Homer Core’s server-side deployment gives platform teams a central governance control plane. It allows users to set org-wide policies ; every team’s SDK enforces them automatically. Tenant isolation keeps each team’s data and audit logs separate. Core provides:</p><ul><li>Central policy management across all teams</li><li>Per-tenant isolation, RBAC, usage dashboards</li><li>SaaS, VPC, or air-gapped deployment</li><li>OpenTelemetry-compatible telemetry and tracing</li></ul><figure><img alt="A robotic hand pointing toward something." src="https://cdn-images-1.medium.com/max/612/1*DniMDF3Xse1GhZQr3Q4udw.jpeg" /></figure><p><strong>Cerone</strong> enables runtime governance for AI agents, managing autonomy, actions, and system boundaries.</p><p><a href="https://pypi.org/project/cerone/">https://pypi.org/project/cerone/</a></p><p>Cerone provides multi-layered protection to monitor and control agent behavior in real time. At a glance, Cerone provides:</p><p><strong>01 — IDENTITY : </strong>Trusted Agent Foundation</p><p>Agents are authenticated with verifiable credentials that establish their intended purpose and scope. Identity forms the foundation of every access decision.</p><p><strong>02 — VALIDATION : </strong>Purpose Alignment</p><p>Agent actions are continuously evaluated to ensure they align with their designated role. Misaligned behavior is detected and prevented before it causes harm.</p><p><strong>03 — GOVERNANCE : </strong>Continuous Oversight</p><p>Trust adapts dynamically based on observed behavior. Access levels adjust automatically to minimize risk while maintaining a complete audit trail.</p><p><strong>04 — TOKEN EXCHANGE : </strong>Delegated Security</p><p>Agents can securely delegate credentials to child agents without exposing sensitive tokens. Cryptographic lineage ensures each delegation is traceable, time-limited, and revocable.</p><p>Both products have been available on Pypi for a while now — while the market is still waking up to the need of Agentic Governance.</p><p>If you are an organization using AI Agents/ Embedded LLM, do reach out. We are looking for design partners for super brisk, focused proofs of concept.</p><p>Additional details can be found on the Homer Websites.</p><p><a href="https://homersemantics.com/">https://homersemantics.com/</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=69dba0bded66" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Looking for POC partners — Runtime AI governance.]]></title>
            <link>https://medium.com/@anantdhavale/looking-for-poc-partners-runtime-ai-governance-1868e979a985?source=rss-b845d0c5e68e------2</link>
            <guid isPermaLink="false">https://medium.com/p/1868e979a985</guid>
            <category><![CDATA[aiinbanks]]></category>
            <category><![CDATA[agents]]></category>
            <category><![CDATA[fintech]]></category>
            <category><![CDATA[generative-ai-compliance]]></category>
            <category><![CDATA[governance]]></category>
            <dc:creator><![CDATA[Anantdhavale]]></dc:creator>
            <pubDate>Sun, 05 Apr 2026 05:02:21 GMT</pubDate>
            <atom:updated>2026-04-05T05:03:02.514Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*0rlFvmR3fvTzv9Ax4q7V0A.png" /></figure><p>“Banking Is the Worst Possible Place to Run Uncontrolled Agents.”: Let us think about this statement for a moment.</p><p>Banks are compliance driven entities. Here’s a laundry list of things they operate under:</p><p>- Model Risk Management (SR 11–7)<br>- GDPR<br>- FINRA<br>- SEC<br>- Basel<br>- SOX<br>- Audit trails, decision traces, action paths.<br>- Explainability / defensibility/ auditability requirements.<br>- Data lineage requirements.<br>- Insider trading controls.<br>- KYC.</p><p>Now imagine a swarm of AI agents, moving across functions, such as :</p><p>-Trading<br>- Research<br>- Investment banking<br>- Retail banking<br>- Risk<br>- Compliance<br>- Customer sensitive data</p><p>Without a semantic runtime, it is hard to answer:</p><p>- Why did the AI make this recommendation?<br>- What data influenced the recommendation?<br>- Was the AI allowed to access that data?<br>- Was there a conflict of interest?<br>- Which regulation applied at that moment?<br>- Is the Agent identified ?<br>- Is the agent intent defined with a bounded context?<br>- And so on.</p><p>This indicates the need to three key components, within the tech ecosystem that surrounds both — embedded AI and AI agents ( these two are different things.)</p><p>- Semantic Control Layer<br>- Context Runtime<br>- Agent Governance</p><p>No doubt some banks can afford to build these internally — others can rely on ready to use solutions such as Homer Core.</p><p>If you are financial service institution — let us talk. We can build these layers for you effortlessly with Homer Core and Cerone. Along with the compliance benefits, Homer Core is also a great AI cost optimizer.</p><p>I am looking to do proofs of concept with financial services companies — if you are a bank/ brokerage house, a credit union, a regulator, or a consulting/ fintech services firm, please do reach out to me.</p><p>info@homersemantics.com</p><p>Visit the website at www.homersemantics.com</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1868e979a985" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[On-Prem AI Document Intelligence System for Privacy and Confidentiality]]></title>
            <link>https://medium.com/@anantdhavale/on-prem-ai-document-intelligence-system-for-privacy-and-confidentiality-76e453bdb045?source=rss-b845d0c5e68e------2</link>
            <guid isPermaLink="false">https://medium.com/p/76e453bdb045</guid>
            <category><![CDATA[document-intelligence]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[generative-ai-tools]]></category>
            <category><![CDATA[compliance]]></category>
            <dc:creator><![CDATA[Anantdhavale]]></dc:creator>
            <pubDate>Wed, 11 Mar 2026 06:18:45 GMT</pubDate>
            <atom:updated>2026-03-11T06:18:45.792Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LPfyVGn9_GGGEoqfgiSLFA.jpeg" /></figure><p>For organizations that do not want their data to be hosted on or routed through public cloud, a private AI document intelligence system can provide better privacy and optimal control.</p><p>Here is a suggested architecture, using Homer Core — the semantic control plane engine. It provides RAG like document/ knowledge intelligence as well as compliance document testing capabilities. Here’s a sample architecture that can be utilized to implement Homer Core on an On-prem infrastructure :</p><p>Internal Documents</p><p>↓</p><p>Homer Core : Document Processing Pipeline</p><p>Embedding Model (local)</p><p>Vector Database</p><p>Semantic Enrichment + Business Context Awareness Infusion</p><p>↓</p><p>Query from Employee</p><p>↓</p><p>Retrieve relevant documents</p><p>↓</p><p>Local LLM generates answer</p><p>This effectivly becomes a privately hosted AI document search system where the entire RAG pipeline : document ingestion, embedding generation, vector storage, and LLM inference runs on internal servers powered by hardware of the organization’s choice.</p><p><strong>What Actually Runs on the Private Server</strong></p><p><strong>1. Homer Core : Document Processing Pipeline</strong></p><p>Homer Core is a system that:</p><ul><li>Ingests documents (PDFs, Word, engineering specs)</li><li>Extracts and cleans text</li><li>Breaks content into smaller chunks, and enriches them semantically</li></ul><p>This prepares documents for AI search. Homer Core utilizes following components :</p><p><strong>Embedding Model</strong></p><p>A local AI model converts text into vector embeddings (numerical representations of meaning). These embeddings allow semantic search, meaning the system understands the intent of queries.</p><p><strong>Vector Database</strong></p><p>Embeddings are stored in a vector database.This database enables extremely fast similarity search across millions of document chunks.</p><p><strong>Local Language Model</strong></p><p>A locally hosted LLM processes the retrieved documents and generates answers. This uses a technique called <strong>Retrieval-Augmented Generation</strong>, where the AI combines:</p><ul><li>retrieved document content</li><li>the user’s question to produce a response.</li></ul><p><strong>2. User Interface</strong></p><p>Employees interact through:</p><ul><li>a search portal</li><li>chatbot-style interface</li><li>enterprise knowledge portal</li></ul><p>If you are a large enterprise (FinTech, Energy, Healthcare, Defense etc.) that cares for its data confidentiality and privacy, this might be the best option for you.</p><p>Reach out to know more : info@homersemantics.com</p><p><a href="http://www.homersemantics.com/">www.homersemantics.com</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=76e453bdb045" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Ceron : The Zero Trust Security Validation,Verification and Governance for AI Agents]]></title>
            <link>https://medium.com/@anantdhavale/ceron-the-zero-trust-security-validation-verification-and-governance-for-ai-agents-8b654f7afde1?source=rss-b845d0c5e68e------2</link>
            <guid isPermaLink="false">https://medium.com/p/8b654f7afde1</guid>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[zero-trust-security]]></category>
            <category><![CDATA[openclaw]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[cyber-security-awareness]]></category>
            <dc:creator><![CDATA[Anantdhavale]]></dc:creator>
            <pubDate>Fri, 20 Feb 2026 04:43:09 GMT</pubDate>
            <atom:updated>2026-02-20T04:43:09.181Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*QErv3mTNXjK_MMPjCFg5qw.png" /></figure><p>As organizations move toward deploying hundreds of AI agents in production, a critical challenge is emerging in terms maintaining <strong>continuous visibility and control </strong>over who each agent is (<strong>identity</strong>), what its purpose is (<strong>intent</strong>), whether it is behaving as expected (<strong>purpose drift</strong>), and when it may need to be decommissioned.</p><p>Ceron addresses this challenge directly. It continuously evaluates agent behavior against defined roles and objectives, detects misalignment, and enforces corrective actions; including automated shutdown, when necessary, before risk materializes. This ensures that agent actions remain aligned with organizational, operational, and security expectations.</p><p>To sum it up again, one of the most significant risks in agentic AI today is <strong>uncontrolled agent spawning,</strong> where agents are created and retired without centralized oversight, traceability, or accountability. Ceron is built to tackle this, through a semantics + business context grounded approach.</p><p>You can find more details here:</p><p><a href="https://aztp.homersemantics.com/">Zero Trust Security for AI Agents</a></p><p>I am looking for potential POC (Proof of Concept) partners, if you are interested, please reach out via the website or here. Email me at info@homersemantics.com</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8b654f7afde1" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Context Aware, Continuous Compliance with Homer Core]]></title>
            <link>https://medium.com/@anantdhavale/context-aware-continuous-compliance-with-homer-semantics-56ebff985561?source=rss-b845d0c5e68e------2</link>
            <guid isPermaLink="false">https://medium.com/p/56ebff985561</guid>
            <category><![CDATA[fintech-startups]]></category>
            <category><![CDATA[venture-capital]]></category>
            <category><![CDATA[compliance]]></category>
            <category><![CDATA[multi-agent]]></category>
            <category><![CDATA[startup-funding]]></category>
            <dc:creator><![CDATA[Anantdhavale]]></dc:creator>
            <pubDate>Mon, 16 Feb 2026 06:51:19 GMT</pubDate>
            <atom:updated>2026-03-12T03:58:47.312Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/608/1*BJ0EQRY3OfBxmIcwapTn_w.jpeg" /></figure><p>Compliance has evolved from checklist-driven, periodic exercises into a global, data-driven, and continuously evolving discipline. Today, regulation operates across jurisdictions, platforms, and time zones. Decisions are evaluated in a much wider context, marking a shift from from verification to interpretation. Most organizations, however are yet to adapt to this change, still relying on systems that were built on older technology, with static flows and cumbersome architecture. There is a need to move toward an ecosystem of wide, distributed models; so as to maintain situational awareness of how policies interact with transactions in real time; how exceptions propagate and also how risk accumulates across functions.</p><p>This is precisely where Homer Core (HC) comes into the spectrum. Not as another compliance application, but as a contextual intelligence layer. Homer is designed around the reality that regulatory obligations do not exist in isolation, rather they are embedded in processes, data flows, decisions, standards, evidences and real-time transactions.</p><p>HC enables organizations to automate compliance, control testing, and regulatory validation using governed AI agents that operate within verified enterprise context. It allows organizations to :</p><ol><li>Continuously validate controls against regulatory and internal obligations</li><li>Detect gaps, conflicts, and inconsistencies before they become findings</li><li>Flag ambiguous or incomplete evidence</li><li>Surface policy and process deviations</li><li>Trace every automated assessment to source requirements</li><li>Maintain auditable decision lineage</li></ol><p>If you are a forward looking enterprise, aiming to introduce an intelligence control and governance layer, do reach out. More details can be found on the website <a href="https://homersemantics.com/">https://homersemantics.com/</a></p><p>Email : info@homersemantics.com</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=56ebff985561" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Stay Away from Vector Databases.]]></title>
            <link>https://medium.com/@anantdhavale/stay-away-from-vector-databases-c6a08947137c?source=rss-b845d0c5e68e------2</link>
            <guid isPermaLink="false">https://medium.com/p/c6a08947137c</guid>
            <category><![CDATA[agentic-rag]]></category>
            <category><![CDATA[generative-ai-tools]]></category>
            <category><![CDATA[database]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[vector-database]]></category>
            <dc:creator><![CDATA[Anantdhavale]]></dc:creator>
            <pubDate>Wed, 10 Dec 2025 19:40:08 GMT</pubDate>
            <atom:updated>2025-12-10T19:40:08.394Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/612/1*oypsUGNva4LdoIJmEndfXQ.jpeg" /></figure><p>The title’s a click-bait, Vector databases are fine, and at times useful; especially if you are a large organization with disposable IT budget. However, like most glittering AI toys, Vector databases too are somewhat redundant; they create a new box (or an index in the DB language) in an already muddled up Technology Ecosystem. Not to mention Vector DBs are expensive.</p><p>There are ways to optimize your existing data storage to make room for vectors. Implementing RAG in an iterative manner allows for sufficient time and room for such modifications.</p><p>Colloquial wisdom teaches us to use what we have before buying anything new. Modification and re-purposing can go a long way if you are looking to optimize technology usage that actually yields value.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c6a08947137c" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>