Enterprise AI, from
prototype to production
Long-form articles on how enterprise AI systems are actually built — control loops, agentic architecture, token economics, and honest analysis of models and benchmarks. Plus shorter field notes on what is changing week to week.
Articles
9 publishedThe LLM Pipeline Problem in AI-SDLC: Error Ossification
Chaining LLMs across the software lifecycle without deterministic gates is an architectural antipattern. Upstream assumptions become downstream axioms, and small ambiguities ossify into production behavior.
From AI Prototype to Production: The Control Loop Enterprise AI Needs
Production agentic AI is not chatbot → gateway → model. It is a managed control layer that validates inputs, measures context quality, evaluates outputs, monitors trends, and escalates when risk patterns change.
The Seven Token Categories Enterprises Should Understand
LLM pricing is not just input and output tokens. Modern AI systems generate tokens across seven categories — many invisible to whoever owns the budget. Where token consumption hides in enterprise AI.
From Visibility to Action: How AI Can Improve Enterprise Command Centers — Part 1
Fleet management, payment reconciliation, airline ops, network centers, and supply chain control towers are all the same pattern. AI can move them from passive visibility to contextual prioritization and action.
Building Agentic AI Systems: Core Concepts
Using AI during development is very different from using AI at runtime. The problem classes suited to runtime AI, the four layers of context, the four kinds of memory, and the observability an agentic system actually needs.
AI Benchmarks I: Gemini 3 Scored 85% on ARC-AGI-2. What That Actually Proves.
Gemini 3 Deep Think set a record on ARC-AGI-2. Is it genuine reasoning progress, or smarter search amplified by compute? What ARC-AGI measures, what it does not, and why benchmark wins don't equal field-readiness.
With Milind Chandramohan
Among the First to Pass: Strategic Insights into the AWS Generative AI Professional Exam
Why the AWS Certified Generative AI Developer – Professional exam is genuinely hard, and what it reveals about AWS's philosophy: GenAI as a system, not a product. A test of architectural maturity, not memorization.
Gemini 3.0 — Hype vs. Reality
Eight of the biggest claims about Gemini 3.0 — the ecosystem moat, the price war, Nano Banana, benchmarks, multimodal, coding, Antigravity, and the knowledge cutoff — measured against benchmarks, pricing, and hands-on testing.
AI Agents: Our Experiments, Key Implementation Techniques, Security Challenges & Observability Solutions
A hands-on field report: four agent implementation patterns (tool-calling, LLM-with-tools, MCP, MCP server), the top security threats, observability with OpenTelemetry, and a Level 1–5 agent maturity model built around a CRM automation agent.
With Milind Chandramohan
Field Notes
Shorter takes on models, releases, and industry moves — published on LinkedIn.
The Claude Hackathon Winners Weren't Engineers
A cardiologist, a construction worker, a personal injury lawyer, a musician — LLMs are collapsing the distance between domain expertise and software creation. But production is a second, harder frontier.
AWS Certified Generative AI Developer — Professional (Early Adopter)
Passed the Professional-level exam as part of the first 5,000 Early Adopters — the most challenging AWS certification taken to date.
What Everyone Gets Wrong About GPT-5
The GPT-5 launch split into 'bust' vs. 'breakthrough' — both miss the pivot from better chatbot to reasoning, efficiency, and scalable agentic infrastructure.
Better Than Claude at Coding? IQuest-Coder-V1
A Chinese open-source model claims to beat Claude 4.5 on SWE-Bench Verified by training on repository evolution — commits, PRs, and review comments — not just code snapshots.
Copilots, Copilot Studio, AI Foundry, and One Million Agents
Microsoft's three-tier agent strategy — M365 Copilot, Copilot Studio, Azure AI Foundry — and what an enterprise agent mesh actually looks like.
Generative AI Leader Certification
Completed Google's Generative AI Leader certification.
Why I Don't Obsess Over Prompts
Prompt engineering is picking better syntax; context engineering is choosing the right data structures. Most production failures are context failures, not prompt failures.
OpenAI DevDay: AppsSDK Flips the Model
Apps no longer connect to ChatGPT — they embed inside it. What that means for data ownership, the customer relationship, and a coming multi-cloud dilemma across model providers.
What the Claude Code Leak Actually Revealed
The leak didn't expose the model — it exposed the harness. The real intellectual property in agentic systems is operational control, not the LLM.
Quiet AI: Stabilizing the Systems That Run the Enterprise
The highest-value enterprise AI isn't rewriting legacy systems — it's a detect (deterministic) → triage → strategy → safety-check (probabilistic) → resolve (deterministic) pattern that keeps them operational.
How Teams Are Actually Starting With Agentic Coding Tools
IDE-first adoption is winning; agent orchestration is still immature. A four-phase adoption path from individual acceleration to selective autonomy.
I Changed My Mind on Running AI Models Inside the Enterprise Boundary
Deployment boundary, data retention, and compliance now matter as much as benchmarks for regulated enterprises — the question isn't just 'which model,' it's 'where is it served.'
The Projected Shape of AI in 2026
AI shifts from feature to infrastructure — physical AI, agent swarms, and industrial-scale AI-native platforms, with compute, energy, and governance as the new design constraints.
OpenAI's Next Act — The Effect of AI at Work
From AGI dreams to a concrete AGI deadline: OpenAI's pivot from philosophy to engineering, and the harder question of what humans do next.
Is Agentic AI Just a Fancy Rule Engine?
On the paper 'State and Memory is All You Need' — FSA + structured memory + LLMs, and how that echoes Drools and Prolog while doing what rule engines couldn't.
Are LLMs Becoming Sentient?
The 'AI blackmails engineer' headlines, the goal-conflict test setups behind them, and why these are cautionary tales about guardrails — not horror stories.