AI Product Engineering without hype
This series collects practical lessons from building AI products in real contexts: real users, real data, real costs, real latency and real teams.
It is not about pretty demos or adding a chatbot to a screen. It is about the less glamorous pieces that make AI hold when it moves from experiment to product.
The articles are dated by the approximate moment when each lesson started in real work, not by the day I wrote the post.
The shorter professional context is in Professional profile and Selected work.
Articles in the series
- 2024-05-09 - Useful RAG is not just embeddings
- 2025-09-19 - Product feedback loops: from loose opinions to decisions
- 2025-10-17 - Pragmatic ranking: heuristics, a cheap LLM and deterministic validation
- 2025-12-05 - Sofia and the real problem with AI tutors
- 2026-02-10 - AI cost is architecture too
- 2026-02-20 - Executable guardrails: living policies, not pretty documents
- 2026-04-24 - Evaluating agents is not asking whether the answer looks good
The thesis
AI does not become useful in a product just because there is a model inside.
It becomes useful when there is:
- well-scoped data;
- retrieval with permissions, fallback and ranking;
- persistent evaluation;
- observable cost and latency;
- executable guardrails;
- product feedback loops;
- deterministic validation before anything reaches the user.
That is the territory I am most interested in.