•AI Product
Evals and Human-in-the-Loop: How I Ship AI Features Without Gambling Trust
A practical playbook for offline evals, online acceptance metrics, HITL boundaries, and tightening autonomy only when the product earns it.
Read Essay
BLOG
Ideas, lessons &
things I'm building
Practical notes from shipping AI products — agents, tokens, context, recovery, evals, and the product decisions that make systems trustworthy after launch.
•AI Systems
Fault Recovery in AI Agent Systems: Assume the Model Will Fail Mid-Flight
Designing durable AI workflows with checkpoints, idempotent tools, Temporal-style recovery, and human takeover — so partial failure does not become silent corruption.
Read Essay
•AI Systems
Context Management for AI Products: Working Memory Is Not the Database
How I design context layers — working memory, canonical state, retrieval, and provenance — so AI features stay useful as products grow.
Read Essay
•AI Systems
Token Optimization Is a Product Constraint, Not a Prompt Trick
How I treat tokens as latency, cost, and quality budget when designing AI features — caching, compression, routing, and saying no to context bloat.
Read Essay
•AI Agents
Production AI Agents Are Product Systems, Not Chat Loops
What I learned shipping agent workflows end to end — from customer jobs and capability limits to tool graphs, state, evals, and post-deploy ownership.
Read Essay