Direct Wines — AI dev infrastructure
AI development infrastructure in the team's existing stack — adoption-first.
Fractional product lead · Direct Wines · concluded March 2026
- PROBLEM
- A 70-person engineering team shipping across multiple platforms, with no AI tooling in place.
- BUILT
- AI dev infrastructure built in the team's existing stack, adoption-first.
- OUTCOME
- 70% faster development · 3× fewer defects · weeks → days new-developer onboarding
- STACK
- Claude API · Claude Code · GitHub Actions · Cursor · VS Code · GitHub Copilot
- STATUS
- handed over
Why it exists
Direct Wines had a 70-person engineering team shipping across multiple platforms with no AI tooling in place. Manual code review ate senior developer time on repetitive feedback, onboarding a new engineer took weeks of reading the codebase, and there was no internal capability to build or integrate AI. They wanted AI to make the development organisation faster — not a pilot, not a deck.
Key decisions
I started by spending two weeks inside the team — shadowing developers, reviewing where work actually got stuck — and picked the highest-leverage automation points across the developer experience instead of the flashiest ones.
Everything was built in the team’s existing stack: no new platform, no parallel tooling the team would abandon in a quarter. Three systems came out of it — AI-powered configuration for the editors the team already used (Cursor, VS Code, GitHub Copilot, Claude Code) with context-aware rules tuned to the codebase; a GitHub Actions PR-review bot with severity-based feedback, where critical issues block the merge and style suggestions don’t; and an AI onboarding guide that lets new developers ask questions about the codebase, architecture decisions, and team conventions.
The result
Development got 70% faster, with 3× fewer defects reaching code review, and new-developer onboarding went from weeks to days. The engagement concluded in March 2026 and the infrastructure was handed over to the team — it runs without me, which was the point. The full breakdown is on fractalocean.ai.