Field notes from real builds, opinions you can disagree with, and the occasional open-source drop. No "ultimate guide to prompt engineering" trash. Just what we've learned this month.
Most "agent" demos collapse the moment they touch a real API. Here's the small set of patterns - sandboxes, allow-lists, approval gates, and deterministic fallback - that we use to ship production agents that don't go sideways.
Why most knowledge-base bots are bad - and the boring infrastructure (chunking, re-ranking, per-tenant data) that makes them good.
RAG · Read →The unglamorous architecture behind an assistant that handles enquiries, bookings, and follow-ups on the world's most-used app. India edition.
VOICE AI · Read →Multimodal models flipped the playbook in 2025. Here's a pipeline you can stand up over a weekend - including the human-review step you'll forget.
DOC INTEL · Read →Every millisecond between user-stop-talking and bot-start-talking matters. The architecture we use to stay under 400ms end-to-end.
VOICE AI · Read →A 200-line library to run regression tests on prompts and agent runs. Catches the "this used to work" failures before they hit production. MIT licensed.
The taxonomy that helps us recommend the right thing to clients - and stops them from buying an "agent platform" they don't need.
AI AGENTS · Read →From cold lead nurture to closed-won onboarding. The flows we've shipped, the ones that flopped, and what makes the difference.
AI AGENTS · Read →Yes, the LLM could do it. We don't let it. A small rant on engineering hygiene.
OPINION · Read →Why your team's most-used software in 2027 won't have a Hacker News thread. And why we love it.
OPS · Read →Hour-by-hour from one of our Forward Deployed Engineers, fresh from a 3-week warehouse deployment. Not glamorous. Effective.
★ FDE · Read →OpenAI vs Anthropic vs Gemini vs open-source. The framework we use to pick - and when we mix.
ADVISORY · Read →Practical writing from the team - what we shipped this week, what broke, what surprised us. About 1,200 words. No motivational quotes, no "5 ways AI will change X."