Insights

Notes from the work itself — how we ship AI features, what we measure, and how we run engagements. Written by the engineers doing it.

Add RAG to an Existing Product Without a Rewrite

Scoping a vertical slice, chunking that respects structure, pgvector vs a hosted vector database, hybrid search with re-ranking, retrieval evals, permissions and a rollout you can switch off.

LLM Routing: Cost, Latency and Fallbacks

Task classes, routing signals that actually matter, a small TypeScript router, fallbacks and retries, prefix and response caching, token budgets, and how to prove a cheaper model is good enough.

Nearshore Software Development in Romania

Why Romania keeps appearing on shortlists, which engagement model fits which problem, how to vet engineering rather than a sales deck, what belongs in the contract, and what a well-run first two weeks look like.

Working on something like this?

Start a project→