“We’ve been told it’s gonna take a lot longer than it should.” That is when the call comes. I take that one project, do the due diligence, and build it in the team’s own stack. Then I teach them to run it and leave.
I’ll tell you in about a day whether I’m the right person. The first conversation is fit, not a free architecture review.
Proof
Two years of design meetings had produced no software. Two weeks in Cursor produced a working prototype — and the demo is what got engineers assigned to it.
Health insurance · Cigna · React/Node on ECS · SQLite · in-app assistantAnother team's invalidation bug went straight through a cache we had already built and saturated a MongoDB we didn't own. The degraded mode we designed in kept bookmarked pages serving, and our own logs named their bug before their team was reachable.
Real estate · Anywhere Real Estate · Lambda mid-tier · CloudFront + S3 cache · shared MongoDBNothing broke. The first workload still took over a year, and almost none of that was the technology — it was the path to prod. Once the path existed, a project crossed in about six weeks. I ran eight through it.
Health insurance · Cigna · PCF → AWS · Terraform · 8 projectsA $700k-a-year renewal was on the table for features the enterprise already owned elsewhere. I forked the query editor and pointed it at the metadata on hand. In prod in two to three months; the vendor is sunsetting.
Health insurance · Cigna · ~1,300 users · DBeaver fork · Collibra · Temporal246 cards merged in one day.
The system I built to run a team of AI coding agents, used until Jarvis replaced it in October 2026 · off the board's own ledger, September 12, 2026, WIP cap 204,884 PRs merged in eight weeks across two codebases, and the rules that got them through now live in a gate, not a prompt.
My current system for running AI coding agents, the successor to Omnigent, rebuilt from scratch after an audit of its first version · off main's merge history, September 5 to October 28, 2026 · Python · React · Temporal
Who I’ve done this for
How a carve runs
I read the estate before I touch it: what’s actually blocking, who owns the gate, what the team already tried. Most stalled projects are not stalled on the technology.
Your languages, your patterns, your design. Architecture judgment transcends the stack; the implementation language is yours, because your team is the one that has to maintain it.
I hand back a working system, not a repo. The skills and the chat lineage go in with it, so the next engineer — and the next agent — can see why. Then I’m gone.
Why a carve
Applied AI
Systems
Writing
Field notes
Next
Start