How Brainbase took an AI agent from 41% to 97% accurate in six steps
AI Engineer · youtube · 2026-11-05
Milind Sunil Shah of Brainbase Labs explains how his team improved a back-office agent tracking Domino's competitors' pizza prices from 41% to 97% accuracy across 11 versions — halving cost per source from $2 to $1 and cutting runtime from 129 to 71 minutes — through instructions, skills, and playbooks rather than a new model.
The six-step framework:
- Kickoff: agree with the client on what "winning" means
- Ontology: define every business term precisely
- Benchmarks & evals: build evals before building
- Ship a narrow first agent, not a general one
- Iterate: score, find failures, fix, re-ship (e.g. missing deals below links, getting stuck zooming into images)
- Hand off the whole loop so the client can keep iterating
A full engineering retrospective with real failure cases — valuable for anyone building production agents.
More from coding & agent
- Teardown: Viral Claude Code Skill Turns Any Machine Into a 3D Simulator You Can Take Apart — huangyun_122 · 2026-11-05
- Evals cheat sheet: a plain-language guide to knowing whether an AI system actually works — blaizedsouza · 2026-11-05
- Dataflare: open-source product analytics that runs in your own Cloudflare account — isnit0 · 2026-11-05
- App Store Connect CLI ships 39 releases in 60 days, hits 8.2k GitHub stars — rudrank · 2026-11-05
- Parcha open-sources Recall, the skill that became their company brain for Claude Code sessions — blaizedsouza · 2026-11-05
- tw93 open-sources Kami: a design system that makes AI-agent documents actually readable — sujingshen · 2026-11-05