AI-Assisted Engineering
This isn't a product I shipped. It's the work of changing how my team and the teams around us actually build with coding agents: making repo context something engineers inherit rather than rediscover, and making the case for it with evidence instead of enthusiasm.
Where the team was
Coding agents were arriving team by team with no shared setup. The same agent would answer correctly in one service and confidently wrong in the next, and nobody could say which was which, because there was no baseline to compare against. Engineers re-explained the same context every session, and people were hand-migrating dashboards one query at a time. The friction was obvious, but the case for fixing it was all anecdote, and "it feels faster" isn't an argument you can take to an engineering org.
How I approached it
I wasn't interested in evangelizing tools, I wanted to move the default. That meant writing repo context down so every engineer inherits it instead of rebuilding it, then proving the value with a controlled evaluation rather than a demo, and publishing the results including the repos where it changed nothing. Once real numbers existed, adoption stopped being a matter of opinion.
What I changed
Three changes to how the work gets done day to day.
Made repo context the default instead of a personal habit
Rolled out AGENTS.md standards across ten backend repos, so the context an engineer used to carry in their head now lives in the repo and applies to every agent session automatically. Then I evaluated the same tasks with and without it. Three of the ten produced a real correctness regression in the baseline that the context file prevented outright. Where both passed, the win was cost: one repo went from a 30-plus turn agent run to a single-turn answer.
Gave the org a shared view of its own health
Co-built a dashboard tracking incidents, mean time to recovery, and deploy frequency, so conversations about how the team is doing start from data rather than impressions. I own the operational-health pillar on the cross-org working group that maintains the framework it feeds.
Turned specialist migration work into something anyone can run
Engineers were hand-converting dashboards out of our legacy analytics tool, query by query. I built a coding-agent skill that does the conversion, which took the work out of the specialist-knowledge category entirely. The platform team running the company-wide migration adopted it as the recommended interim path and assigned a pilot team to stress-test it on harder dashboards.
Made the case outside my own team
Gave the engineering half of "AI at Lyft: From Data Foundations to Shipping Faster," a sold-out Lyft Engineering session at Toronto Tech Week 2026 with just over 100 attendees. Lyft's Head of Corporate Data & Analytics covered the data foundations; I covered delivery: how a continent-scale Lyft and FREENOW integration that should have needed a year of ramp-up didn't, and where Cursor and Claude concretely changed how we plan, write, and review code. We closed on a joint panel tying the two halves together, which is the honest version of the argument. Agent-assisted delivery only compounds when the foundations underneath it hold.
Toronto Tech Week 2026 event page ↗
How it spread
Ten backend repos now inherit shared agent context by default, no per-engineer setup
The evaluation was published openly, including the repos where it made no difference, so adoption ran on evidence
The migration skill became the recommended interim path for a company-wide tooling migration, with a pilot team assigned to it
Seats on cross-org working groups for operational excellence and bug-triage automation
A public account of the European build, given to a sold-out room during Toronto Tech Week
Next
Search Optimization
Streamed multi-provider search results over WebSocket so users saw imagery as soon as the fastest provider responded.
Learn more →