Daniel R. Cooley

Daniel R. Cooley

AI engineer · Los Angeles · Remote

I design, ship, and operate production multi-agent LLM systems end to end: orchestration, model routing, structured outputs, the evals that prove it works, and the infrastructure it runs on. Everything below is live or measured, and I built it.

Email me LinkedIn See the work

Shipped work

StoreXray

Live multi-agent SaaS · sole builder · 2024–present

Paste a Shopify store URL, pay $49, and an orchestrator runs seven specialist Claude agents (scrape, UX, speed, funnel, accessibility, AI visibility, roadmap) that return a 30+ page repair plan. Nobody touches it between payment and delivery.

  • Each agent is pinned to the cheapest Claude model that holds quality for its task (Haiku, Sonnet, Opus). About $1.20–1.40 of API cost per audit and roughly a 96% gross margin.
  • Schema-validated outputs (Pydantic), retries typed to rate limits, timeouts and 5xx errors, and per-agent fallbacks, so one bad agent never sinks a run.
  • A deterministic quality gate holds degraded runs, and an Opus reviewer checks every report against the raw evidence before it ships.
  • A measured WCAG 2.2 A/AA accessibility scan, also sold on its own for $29. axe-core runs in a real browser on five pages through checkout, the agent can only explain rules the scan actually failed, and the gate holds any claim that a store is compliant.
StoreXray landing page: the $49 audit before the agency, with a store URL field and a Run Audit button.
Report table listing the seven agents (Scrape, UX, Speed, Funnel, AI Visibility, Accessibility, Roadmap), what each analyzed, and the evidence each captured.
Every report shows which agent found what, and the evidence behind it.
Executive summary of a sample audit: the three issues hurting revenue, each with severity and estimated impact, and a 30-day priority list.
Executive summary from a real run, store name redacted.
  • Python
  • Anthropic API
  • Pydantic
  • Playwright
  • axe-core
  • Cloudflare Workers + R2
  • Stripe
  • Linux VPS
getstorexray.comAccessibility scanFull sample report (PDF)

LabelCheck

AI label-verification prototype · take-home for a federal AI role · 2026

Checks alcohol label artwork against the application data the way a compliance agent would: brand, ABV, and the government warning across front and back labels, with results in seconds and batch upload that pairs hundreds of images to their applications. Built for reviewers who aren't technical, so the whole flow is one screen.

Real time, no edits: five applications and ten label images checked in about 14 seconds.
Evidence panel: each check shows what the application says, what the label says with a crop of the label text, the matching federal regulation, and the label images with the read text highlighted.
Each check shows the application value, the text read from the label, a crop of where it was found, and the regulation behind it.
  • Python
  • FastAPI
  • RapidOCR (ONNX)
  • Azure Container Apps
  • GitHub Actions CI
labelcheck.devSource on GitHub

MemeReverse

Mobile app on the App Store and Google Play · sole builder

Point it at a meme and it finds the original template: CLIP embeddings in pgvector to match the source image, Google Vision OCR to find the overlaid text, and LaMa inpainting to erase it. When nothing in its template set is a close enough match, it falls back to a web search through Google Lens and then Google Vision, with results cached and the paid lookups capped per day and per month.

Reverse search screen showing an exact match for the Grumpy Cat template and similar templates below it.
Remove text screen showing a meme image with its caption erased by AI inpainting.
Caption editor with font, color, size and outline controls.
  • React Native
  • Expo / EAS
  • Supabase + pgvector
  • CLIP
  • Google Lens + Vision
App StoreGoogle Play

Parity

LLM evaluation research · 2026

One model name on OpenRouter is served by many providers. Parity pins each provider endpoint and runs the same test suite against all of them (exact answers, JSON-schema output, tool calls, long-context retrieval, latency, and billed cost) to see whether they really behave the same.

  • Found that Google Vertex endpoints for Claude Sonnet 5 returned HTTP 404 on every JSON-schema request (24 of 24 across three sweeps), reproduced it, and reported it to OpenRouter.
  • Every number in the write-up is re-derived from raw run logs, and the write-up was reviewed by a second model three times before sending.
  • Python
  • httpx
  • JSON Schema
  • pytest
  • LLM-as-judge
Ask for the findings

Arktista

Construction AI platform · two-person team

Takes a home remodel from a phone scan or blueprint to a permit-ready packet, with a licensed professional signing off. Constraint-solved floor plans, a permit-plan engine that cites code sections rather than guessing, and production auth and billing.

  • TypeScript
  • React
  • Fastify
  • PostgreSQL
  • Auth0
  • 441 tests

Currently in development

How I build

Evals before agents

I write the test set and the scorer first, then build until the number moves. I don't trust a result I can't reproduce.

Evidence, not vibes

Findings point to what they're based on. Reviewers check output against the raw input, not against how confident it sounds.

Cost and failure are design inputs

Model routing, budgets, typed retries and fallbacks are in the first version, not bolted on after the bill arrives.

AI tools as a power user

I build with Claude Code and Codex daily, including second-model code reviews on every major change.

Background

Before software I was a union commercial HVAC installer and foreman, working on precision sites including NASA's Jet Propulsion Laboratory and Caltech clean-room labs. That's where I learned to read a spec, run a crew, and finish the job. Since January 2024 I've built AI products full time.