I build AI agent systems that actually ship.

Fifteen years of Rails, now spent designing multi-agent architectures, the tooling that makes them trustworthy, and the prototypes that close the meeting. I ship in public and measure everything I build.

What I do ↓ GitHub — 70 repos · 376★ ↗ LinkedIn ↗

01What I do

Five things I'm good at

Design agent architectures that stay cheap

Triage before dispatch — single-digit-dollar agent teams where a naive swarm costs hundreds. roundhouse · dev-loop

Go deep on Rails, fast

Fifteen years auditing, refactoring and rescuing codebases, from two to two hundred engineers. rails-audit · boorails

Build tooling that makes agents trustworthy

Skills that force the model to run and return a receipt with a file path — not narrate. boorails · prompt-refiner

Ship prototypes that close the meeting

Production-feel, zero build, no lorem ipsum. Every button goes somewhere. claude-prototype

Instrument everything, publish the failures

I put the numbers — including the ugly ones — where anyone can check them. the proof below

02Proof, not promises

Capability 01, measured

Two multi-agent architectures, the same ten Rails tasks, interleaved A/B, real money on the meter. The finding: 69% of a multi-agent system's bill is agents talking to each other — not writing code. The work wasn't the bottleneck. The meeting was.

69%

of spend was coordination. Peaked at 85.5%.

12.2×

cheaper like for like. And 2.1× faster.

$9.10

for a swarm to fix a one-character typo.

The full ledger — 10 tasks, $144.72 open ↓

Task	Swarm	Mine	Ratio
Fix a flash typo	$9.10	$0.36	25.6×
Missing translation	$11.90	$0.35	33.6×
Unique slug + index	$13.19	$0.70	18.9×
Post.recent scope	$10.01	$0.61	16.5×
Extract a service	$19.78	$1.68	11.8×
User email specs	$6.71	$0.48	14.1×
Comment resource	$21.93	$1.52	14.4×
Publish notification job	$12.95	$1.83	7.1×
Admin API + auth + rate limit	$18.17	$2.65	6.9×
Async report refactor	$4.55 ✗	$6.25	no code
Totals (9 completed)	$123.74	$10.18	12.2×

The last row failed on both sides — the swarm burned its whole budget writing a spec and shipped no code, yet the harness still reported tests_pass: true. A green check can be a lie. Raw data + harness →

03What I shipped

Six tools, all public

roundhouse

Rails specialist team, triage-first. 7–34× cheaper.

rails-audit

18-dimension read-only stability audit.

boorails

7 lifecycle skills that run, not narrate.

dev-loop

6-phase plan → build → adversarial review.

prompt-refiner

Turns loose intent into an executable spec.

claude-prototype

Clickable production-feel demos, zero build.

$ claude plugin marketplace add kurenn/marketplace

04Where I was wrong

Four that didn't work

Shipped three Claude Code hooks that never ran, for two releases. A silent no-op is indistinguishable from success at a glance.

My benchmark lied to me — headless Chrome's 500px floor faked the mobile clipping. I published the correction on top of my own results.

My liveness probe kill-looped nine healthy agents. The monitoring was the incident.

"Calibration on N=2 isn't calibration" — so I wrote my own coverage gaps into the README before anyone else could.

05The record

First Rails app in 2010. Wrote APIs on Rails in 2014. Today I'm CTO and co-founder of Coba (YC S23), a board member at Icalia Labs, and a contributor to Codeando México. Seventy public repos, 376 stars, roughly ten commits a day. The findings aren't Rails findings — "69% of your bill is coordination" isn't a fact about Ruby. It just happens to be measurable in the stack I know cold.

06Work with me

Want this pointed at your codebase?

I take one sprint client a month — one week, one working AI application. Tell me what you're building and I'll reply within a day.