Speed of execution here is wild — feels like we’re moving from “tools assisting work” to “tools compressing entire workflows.” Curious how you’re thinking about scaling and maintainability as complexity grows.
Katy Sander and I joined forces to trap a demon and bind it to the Mac notch. He wasn’t happy about it. But he’s just as helpful as the other agents, and a lot funnier.
Meet Notchling. Hand-drawn, then vectorized and rigged in Rive. All animations, state machines, and timelines were built using the CLI, along with the ElevenLabs and Claude integrations that bring him to life in the final Swift package.
I build AI that’s measured before it ships. Now open for freelance work.
Most AI features fail quietly. No one measured them, and the numbers come straight from the model. My work fixes both.
What I’ve built:
→ A multi-agent tax copilot for UAE–US businesses, live in production. The LLM reads the documents and cites the rule, and a deterministic engine does every calculation. 89% answer correctness on a human-reviewed eval.
→ An evaluation harness for legal AI. It benchmarked three frontier models on contract clauses (F1 0.88–0.90) and found a failure mode all three share: they cut off the legal condition that actually matters.
→ A brain MRI research platform with 3D lesion segmentation, sealed audit trails and FHIR export.
→ A pricing tool for metals traders that turns offer sheets (Excel, PDF, photos, voice) into priced bids.
→ Voice agents: a voice-to-clinical-note scribe, and a browser agent you control by speaking.
How I can help:
• AI agents and RAG systems, with evals built in
• Reliability audits for an AI feature you already shipped
• Document extraction from PDFs, sheets and scans
• Voice agents for calls and intake
• Healthcare AI workflows
If you have an AI feature that’s unreliable, or one you haven’t built yet, send me a message. First 3 clients get a priority start this month.