<?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 Pijush Chatterjee on Medium]]></title>
        <description><![CDATA[Stories by Pijush Chatterjee on Medium]]></description>
        <link>https://medium.com/@pijushchatterjee?source=rss-d09ab48244a7------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*FQ0ra-Exu9lwS_T5.</url>
            <title>Stories by Pijush Chatterjee on Medium</title>
            <link>https://medium.com/@pijushchatterjee?source=rss-d09ab48244a7------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 07:53:29 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@pijushchatterjee/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[Digitizing Enterprise Architecture: slideware to software]]></title>
            <link>https://medium.com/@pijushchatterjee/digitizing-enterprise-architecture-slideware-to-software-a58f00deeeb0?source=rss-d09ab48244a7------2</link>
            <guid isPermaLink="false">https://medium.com/p/a58f00deeeb0</guid>
            <category><![CDATA[ai-engineering]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[enterprise-architecture]]></category>
            <category><![CDATA[digital-transformation]]></category>
            <dc:creator><![CDATA[Pijush Chatterjee]]></dc:creator>
            <pubDate>Wed, 11 Nov 2026 22:54:09 GMT</pubDate>
            <atom:updated>2026-11-11T22:54:09.798Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>Digitizing Enterprise Architecture: slideware to software</strong></p><p>AI is changing enterprise architecture from the inside out. Code is now fast and cheap to produce, so what makes a rollout successful is clear decisions and guardrails.So EA’s job shifts from reviewing designs at the end to shaping delivery from the start.This includes recording decisions as versioned artifacts, encoding standards as executable policy, and handing over build-ready packages to engineers and AI agents directly. Architecture can exist in the delivery system instead of documents. Enterprise Architect function will be measured by the speed, safety and value of what the enterprise ships.</p><h3>The pinwheel</h3><p>Six blades turning on one hub, each producing a digital output</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*bZbx15dlbhm7DxG1jUZCFw.png" /></figure><h3>From legacy EA to digital EA</h3><p>Here’s how I see the shift</p><p><strong>Diagrams:</strong> hand-drawn → one model, views generated</p><p><strong>Decisions:</strong> buried in minutes → versioned architecture decision records (ADRs)</p><p><strong>Standards:</strong> PDFs → guardrails as code</p><p><strong>Governance:</strong> review boards → automated checks; human review saved for irreversible decisions</p><p><strong>Handover:</strong> meeting/docs → a versioned pack</p><h3>The digital outputs of EA</h3><p>For each blade, start with one input, one prompt and one output:</p><p><strong>1.Use-case spec.</strong> In: request, notes, KPIs. Prompt: <em>“Fill the intake schema; flag gaps.”</em> Out: scored spec.</p><p><strong>2. Decision log.</strong> In: spec, constraints. Prompt: <em>“Compare options; draft ADRs.”</em> Out: ADRs.</p><p><strong>3. Architecture model.</strong> In: spec, ADRs, landscape. Prompt: <em>“Model it as C4 with data and trust boundaries.”</em> Out: model and generated views.</p><p><strong>4. Guardrails.</strong> In: invariants, risk tier. Prompt: <em>“Turn each invariant into a policy or test.”</em> Out: policy as code, fitness functions.</p><p><strong>5. Handover pack.</strong> In: artifacts 1–4, contracts. Prompt: <em>“Package thin slices with tests and traceability.”</em> Out: build-ready v1.0</p><p><strong>6. Learning record.</strong> In: run metrics, incidents. Prompt: <em>“Compare to baseline; propose rule updates.”</em> Out: revised patterns.</p><h3>Measuring Enterprise intelligent maturity</h3><p>The following KPIs could make maturity measurable continuously:</p><p><strong>Decision latency:</strong> intake to approved ADR (weeks to days)</p><p><strong>Guardrail automation:</strong> standards enforced as code (prose to executable)</p><p><strong>Reuse ratio:</strong> solutions built from patterns (bespoke to assembled)</p><p><strong>Handover first-pass yield:</strong> packs built without rework (rework to right first time)</p><p><strong>AI-assisted delivery:</strong> build work by agents within guardrails (experiments to govern throughput)</p><p><strong>Value realization:</strong> initiatives meeting KPI targets (delivered to proven)</p><h3>Closing thought</h3><p>Digitizing the Enterprise Architecture function can provide clarity and clean execution. Intent is captured as a schema, decisions as versioned records, standards as code and handovers as build-ready packs. Enterprise architects will lead through paved roads rather than review gates and take ownership of the front of AI-assisted delivery.</p><h3>Disclaimer</h3><h3>Important Note: The perspectives and insights shared in this post are exclusively my own. They do not, in any way, reflect the views, policies, or endorsements of any current or former employer. This content is for general informational purposes only and should not be construed as professional advice.</h3><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a58f00deeeb0" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[From Data Warehouses to LLMs: Building Reliable Context for Enterprise AI]]></title>
            <link>https://medium.com/@pijushchatterjee/from-data-warehouses-to-llms-building-reliable-context-for-enterprise-ai-64c24ac3579b?source=rss-d09ab48244a7------2</link>
            <guid isPermaLink="false">https://medium.com/p/64c24ac3579b</guid>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[generative-ai-strategy]]></category>
            <category><![CDATA[context-engineering]]></category>
            <category><![CDATA[enterprise-ai-strategy]]></category>
            <category><![CDATA[modern-data-architecture]]></category>
            <dc:creator><![CDATA[Pijush Chatterjee]]></dc:creator>
            <pubDate>Thu, 29 Jan 2026 14:27:39 GMT</pubDate>
            <atom:updated>2026-01-29T14:27:39.302Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/100/0*G0qD6T-Jvdwa_zXO" /></figure><p>January 25, 2026</p><p>Traditional large enterprises face a critical question: How do we integrate LLMs with our existing tech stack to generate real business value? I have shared here some of personal thoughts</p><p><strong>Looking back:</strong></p><p>Enterprise analytics has been built on SQL for almost three decades. Our Data Warehouses and Lakehouses excel at massively parallel processing — filtering, sorting, and grouping data to surface insights. But this architecture has fundamental limitations:</p><p><strong>The Superuser Bottleneck</strong>: Complex SQL queries require deep knowledge of tables, metadata, and business logic. Only a handful of power users possess this expertise. Everyone else relies on pre-built dashboards with limited flexibility.</p><p><strong>The Insight Gap</strong>: SQL retrieves data, but not meaning. Analysts still spend significant time interpreting results in spreadsheets before delivering business insights. The spreadsheet remains king because it bridges the gap between raw data and actionable intelligence.</p><p><strong>Where GenAI Changes the Game: Contextual Analysis</strong></p><p>LLMs with advanced reasoning capabilities can assume the analyst role — crunching data, performing analysis, and generating visualizations that human analysts verify rather than create from scratch.</p><p>This is Contextual Analytics: Domain-specific business data + AI-powered analysis = Intelligence</p><p>Context is the key differentiator. It’s a logical grouping of data — Snowflake tables, Iceberg tables, Google Analytics, third-party feeds, podcasts, videos, web search results — that forms the foundation for analysis.</p><p>For rest of the blog, will focus on enterprise’s own data stack</p><p>The Implementation Challenge</p><ol><li><strong>RAG Complexity</strong>: Retrieval-Augmented Generation implementations struggle to maintain high-quality, relevant context at scale.</li><li><strong>SQL Generation Risk</strong>: Relying on LLMs to translate natural language into correct SQL and navigate complex data models introduces significant reliability concerns.</li></ol><p>A Better Approach: Learning from Data Warehouse Design</p><p>When we built Data Warehouses, we started with business questions, not technology:</p><ul><li>Map domains to data entities</li><li>Identify cross-subject relationships</li><li>Discover new questions enabled by integrated data</li></ul><p>Apply the same discipline to AI: Context + Instruction Set + Model = Response</p><p>This formula becomes your use case foundation for each business domain.</p><p><strong>The Framework: Build Reliable Context Fast</strong></p><p>Step 1: <strong>Mine Your Production Queries</strong> Extract SQLs from Data Warehouse metadata tables (last 6 months). These queries represent proven, production-tested code that already navigates your data models successfully. Include frequently-run Python queries from Lakehouses.</p><p>Step 2: <strong>Add Semantic Metadata Tag each SQL/Python</strong> script with metadata describing its output and business context/domain. This creates a knowledge graph of your analytical patterns.</p><p>Step 3: <strong>Build Context Files</strong> Create structured context (markdown works well) where each segment represents query output collected from above as applicable. This gives LLMs highly reliable, pre-validated data. Bonus: Markdown provides dense vector embeddings for tokenization.</p><p>Step 4: <strong>Map Analyst Workflows </strong>Pair business SMEs with technical experts to document the “actions/task sequences” analysts perform on the above data context to generate insights. These workflows become your LLM instruction sets.</p><p><strong>Model Selection Strategy:</strong> <em>Start with the most capable large models to establish acceptable output quality, then benchmark smaller models. Match model capability to task complexity as this affects latency, cost ..etc dynamics.</em></p><p>Step 5: <strong>Establish Quality Baselines Capture “golden responses”</strong> for each instruction type. These become your evaluation benchmarks. In the beginning start with building a checkpointing application with direct human oversight for evaluation and feedback loop</p><p>Step 6: <strong>Technical and Business Validation</strong> Before funding projects, evaluate: model regional availability, RPM/TPM throughput, latency requirements, projected costs, and expected ROI.</p><p><strong>Intelligence Pods: Where Innovation Happens</strong></p><p>Create cross-functional Intelligence Pods for each Line of Business. Business SMEs and tech teams collaborate on “prompt-based data crunching and context-level multi-task analysis” experiments.</p><p>This incubation approach reveals what high-reasoning LLMs can actually solve in your environment, enabling informed prioritization with clearly defined success criteria.</p><p><strong>The Control Paradox</strong></p><p>Here’s the critical question enterprises must answer: What does “AI Agent” mean in your organization?</p><p>Traditional enterprise systems maintain strict application control. When steps fail, runbooks provide restart instructions. Data checkpointing enables recovery. Full telemetry monitoring precedes deployment. We know exactly what’s happening at every point in the workflow.</p><p>Autonomous agents operate differently. They decide which MCP servers to access, which tools to invoke, which functions to call — following their own chain-of-thought reasoning.</p><p><strong>Control shifts from the application layer to the agent.</strong></p><p>This isn’t a technical problem — it’s an architectural decision. How much orchestration control do you retain versus delegate?</p><p>The answer determines whether your agentic workflows can become production-deployable applications. In enterprises, we only scale what we can reliably support.</p><p><strong>Key Takeaway</strong></p><p>Context engineering isn’t about replacing your data infrastructure — it’s about creating a layer that makes your existing analytical assets accessible to LLMs in a structured, reliable way. Start with proven queries, build quality context, establish clear controls, and let cross-functional teams discover value through experimentation.</p><p>The enterprises that succeed won’t be the ones with the most sophisticated AI — they’ll be the ones that systematically bridge the gap between their data reality and AI capability to operate at a higher intelligent state.</p><p>Human Authored, Happy reading !</p><h3>Disclaimer</h3><h3>Important Note: The perspectives and insights shared in this post are exclusively my own. They do not, in any way, reflect the views, policies, or endorsements of any current or former employer. This content is for general informational purposes only and should not be construed as professional advice.</h3><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=64c24ac3579b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Disclaimer]]></title>
            <link>https://medium.com/@pijushchatterjee/disclaimer-701c1b2f7e16?source=rss-d09ab48244a7------2</link>
            <guid isPermaLink="false">https://medium.com/p/701c1b2f7e16</guid>
            <category><![CDATA[enterprise-ai]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <dc:creator><![CDATA[Pijush Chatterjee]]></dc:creator>
            <pubDate>Sun, 15 Jun 2025 23:14:49 GMT</pubDate>
            <atom:updated>2025-06-16T14:48:48.468Z</atom:updated>
            <content:encoded><![CDATA[<h4>Important Note: The perspectives and insights shared in this blog are exclusively my own. They do not, in any way, reflect the views, policies, or endorsements of any current or former employer. This content is for general informational purposes only and should not be construed as professional advice.</h4><h3>Building the Intelligent Enterprise: Your Blueprint for AI adoption</h3><p>For 28 years, I’ve had a front-row seat to technology’s relentless evolution. From the sturdy foundations of client-server systems to the agile dance of data lakes and the early whispers of machine learning, each paradigm shift, brilliantly enabled by Moore’s Law, brought a tangible, quantifiable payoff for enterprises. Storage shrank, compute power soared, and the cloud reshaped our digital landscapes. The ROI was clear, the path, well-trodden.</p><p>But now, we stand at the precipice of a new era — the Age of Intelligence — and it feels different. The immediate, clear-cut value proposition of AI, particularly Generative AI, seems elusive for many. Why the hesitation? And more importantly, how do we truly unlock the profound transformation that AI promises? If the Internet defined our past 25 years, AI is poised to define our next. This is my take on how enterprises can not just adopt, but embody this intelligent future.</p><p><strong>Beyond Efficiency: Crafting the “Intelligent State”</strong></p><p>Forget simply optimizing existing processes. AI’s true power lies in elevating the very <em>experience</em> of how we operate, interact, and deliver value. Consider the omnipresent influence of AI, enhancing our senses and shaping our expectations across every facet of life — from entertainment to social media, cities to individual enterprises. The systems we build will be direct reflections of these heightened expectations and the services we truly value.</p><p>At its core, AI will elevate enterprises to a higher state of “intelligence maturity.” It’s not just about tweaking workflows; it’s about fundamentally transforming business processes and our day-to-day work. The journey into GenAI begins not with a technological mandate, but with a deep dive into your business goals. Identify the pivotal shifts needed in current processes, or envision entirely new ones, that can be meticulously tracked and measured to move your business’s needle.</p><p>Lets look at an example use case:</p><h3>Redefining Customer Journeys: The AI-Powered Returns Experience</h3><p>Forget the traditional friction of product returns. In the Age of Intelligence, we’re crafting new, seamless customer experiences.</p><p>The Old Way (A Chore): A customer sends an email, gets a call, explains the reason, receives a shipping label, ships, waits for receipt and inspection, then money issued. Each step is a potential point of frustration.</p><p>The New Way (A Delight): The customer simply interacts with an “Assistant Agent” to return a shipment. This intelligent agent automatically tracks purchases, confirms the item, prompts the customer to take a pic for real-time validation against stock images, generates a shipping label, shares it instantly, issues the return payment immediately, saves transaction data, and sends a confirmation email. This transforms a potential pain point into a moment of brand delight.</p><h3>The Blueprint for the “Intelligent Agent”: Orchestrating Autonomous Workflows</h3><p>Building this seamless “Return Concierge” isn’t about monolithic programs; it’s about architecting an intelligent, composable ecosystem of AI agents and functions. This involves:</p><blockquote><strong>Deconstructing into Intelligent Tasks</strong>: Breaking down the process into discrete tasks, each assigned an “intelligence grade” (Simple, Medium, High-reasoning) based on the complexity required (e.g., “Generate Shipping Label” vs. “Validate Item Condition”).</blockquote><blockquote><strong>Identifying Data and Tooling Dependencies</strong>: For each graded task, pinpointing necessary data sources (e.g., purchase history, inventory) and specific tools/functions (e.g., shipping carrier API, payment processing API).</blockquote><blockquote><strong>Orchestrating the Composable AI Agentic Workflow Stack</strong>: A central Orchestrator Agent (often a sophisticated LLM) dynamically guides the interaction, invoking specialized sub-agents or functions as needed. Examples include a User Interface Agent, Contextual Memory Agent, Purchase Verification Agent, Visual Inspection Agent (for photo validation), Logistics &amp; Fulfillment Agent, and Payment Processing Agent</blockquote><blockquote><strong>Fine-Tuning Agentic Behavior</strong>: This involves strategic LLM model choices based on task complexity and precise prompt engineering to guide agent behavior, ensure accuracy, and control output formats. Robust guardrails ensure outputs are safe and adhere to business rules.</blockquote><p>These composable agents, exposed via APIs, integrate seamlessly into existing applications. This agentic approach offers unparalleled flexibility, scalability, and the rapid creation of truly intelligent, human-like experiences.</p><p><strong>The New Rules of Software Engineering in the GenAI Era:</strong></p><p>This isn’t just a technological upgrade; it’s a fundamental shift in how we approach software engineering.</p><p><strong>1. Business &amp; Tech, Hand-in-Hand:</strong> GenAI projects demand integrated teams where business stakeholders and technologists collaborate from inception to deployment. The “experience” is designed together.</p><p><strong>2. UX Meets Analytics at the Drawing Board:</strong> GenAI collapses the traditional divide between front-end user experience and back-end data analysis. Defining the desired “experience” becomes the paramount design principle.</p><p><strong>3. Vibe Coding: The Agile Evolution:</strong> Gone are the days of long development cycles. Developers design, and AI tools rapidly code and test. This is agile working on steroids, where the “vibe” of creation is immediate and iterative.</p><p><strong>4. FinOps Granularity: Every Token Counts:</strong> Cost analysis moves from application-level budgeting to granular tracking per LLM call, based on input and output tokens. This new “per API invocation” cost model demands a fresh perspective on financial operations.</p><p>The transformation isn’t coming; it’s here. And the enterprises that embrace this shift, understanding the profound difference between simply adopting AI and truly building an “Intelligent State,” will be the ones that define the next chapter of digital evolution.</p><h3>Disclaimer</h3><h4>Important Note: The perspectives and insights shared in this blog are exclusively my own. They do not, in any way, reflect the views, policies, or endorsements of any current or former employer. This content is for general informational purposes only and should not be construed as professional advice.</h4><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=701c1b2f7e16" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>