Conversations on data, AI, and the systems behind them.
Long-form interviews with engineers, researchers, and operators working at the production edge of AI.
Currently · taking a few client engagements
AI consulting for teams shipping real ML and generative-AI systems.
I'm Muhammad Afzaal — 20+ years building enterprise software, the last five on production LLM and data platforms. I help teams move from prototype to production without losing the plot.
- 20+ years shipping enterprise software
- 5+ years on production LLM systems
- D365 F&O · MCP · RAG · Data platforms
Recent engagements that shipped to production.
Three projects that show what I bring: pragmatic system design, a bias toward measurable outcomes, and code that survives in real environments.
Enterprise ERP team
Built the d365fo-client Python library + MCP server: 29 tools, OData query, FTS5 metadata search. AI assistants can now operate D365 F&O via natural language.
- Python
- MCP
- OData
- FTS5
Generative search platform
Production RAG with chunking, hybrid retrieval, and reranking. Cost dropped 60% on the same answer quality after evaluation-driven optimization.
- LangGraph
- Ragas
- pgvector
Data governance initiative
Mapped 6 OSS data-governance frameworks against NIST AI RMF; published the comparison + architecture guide that anchors team decisions.
- Governance
- OSS
- NIST AI RMF
Notes from production AI.
Lessons from real systems — RAG cost economics, MCP integration, governance frameworks, evaluation pipelines that don't lie.
Why Your LLM is a Stochastic Process (And Why Temperature=0 Doesn't Save You)
A deep dive into why Large Language Models are auto-regressive Markov chains, how GPU floating-point non-associativity and FlashAttention break determinism, and why agent pipelines behave like controlled stochastic systems.
Read
The ERP Copilot Security Dilemma: Dynamic Row-Level Security and Identity Delegation in LLMs
Naive Service Principal access exposes sensitive ERP data to LLM context windows. Learn how to architect zero-trust AI agents using OAuth2 On-Behalf-Of (OBO) token exchange and database-enforced Row-Level Security (RLS).
Read
Beyond the Prompt: Context Engineering Patterns for Complex Enterprise APIs
Enterprise APIs are too massive for LLM prompts. Discover context engineering patterns like JIT schema pruning, semantic routing, and session compression to build efficient agents.
Read
The Quiet Genius Who Made Randomness Calculable
Brownian motion had been observed, priced, explained, and made rigorous. Kiyosi Itô did something stranger: he built a calculus for paths too jagged to differentiate.
Read
A drop-in chat widget you can deploy in minutes.
Open-source assistant with LangGraph orchestration and a clean theme system. The same widget that powers the chat on this site.
Have a hard AI problem?
I work with a small number of teams each quarter — usually on production LLM pipelines, MCP integrations with enterprise systems, or evaluation frameworks that survive contact with real data.