AI Content Platform
TypeScriptBunHonoAWS BedrockPostgresLLMs
team.delivery-process
From random daily requests to one calm path from idea to shipped.
- 01Everything new enters through triage with a short intake form. No request skips the line.
- 02A weekly session with the client turns each request into a decision: build now, discover first, or decline.
- 03Cheap-to-undo work is prototyped in days. Expensive-to-undo decisions are written down before anyone builds.
Overview
At Drizzle, I lead a team of five senior engineers building an LLM platform for a client. It turns large volumes of source knowledge into reviewed, ready-to-use material, with every claim traceable to where it came from. I joined as technical lead and quickly took on delivery and product requirements too, so a big part of the job is reshaping how the team works for an AI-first pace, where shipping a prototype is often faster than discussing a spec. I stay hands-on in architecture, model routing, and code review.
Highlights
- Replaced ad-hoc daily requests with one intake path: everything goes through triage, and a weekly session with the client turns each item into a clear decision.
- Introduced one rule for how deep requirements go: how expensive is it to be wrong? Cheap changes get a prototype in a day or two; data model, security, and integration decisions get written down first.
- Moved senior engineers from detailed tickets to outcome-level ownership, with a named owner for each product area.
- Merged duplicated planning boards into one team workspace without losing history, and tied long-term goals to day-to-day work through a single MVP initiative.
- Made delivery AI-native: shared agent rules, adversarial sub-agent code reviews, and small stacked PRs.
- Hands-on work includes source ingestion with traceable references, model failover with usage and cost tracking, a security hardening pass, and faster CI.