An automated pipeline that finds cow walking footage, pre-labels it with a vision model, and hands our vet a keyboard-first web inbox for review. Built at speed with Cursor and Claude Code.
The short version
Review speed
1 keystroke per clip
Pipeline
One command
Built with
Cursor + Claude Code
The pipeline, drawn
HerdSense builds camera-based lameness detection for dairy farms, and a lameness model is only as good as its ground truth. Ground truth needs a vet. Our veterinary reviewer's time is the scarcest resource in the company, so every hour she spends hunting for files or fighting a clunky tool is an hour of labelled data we never get.
One command runs the whole flow: discover new footage, deduplicate it against everything we have already seen, quality-gate it, extract clips of individual cows walking, pre-label each clip with a vision model, and publish the batch for review. The pre-labels mean the vet starts from a draft rather than a blank page.
Review happens in a web inbox built for speed. Clips play one after another, and a single keystroke (1 to 5 on the Sprecher locomotion scale, the standard vets already use) scores the clip, saves it, and advances to the next. A binary lame-or-healthy label is derived from the score, and a weekly report keeps the team honest about throughput.
Reviewed labels flow straight back into model retraining, so every batch the vet scores makes the next model better.
I built the pipeline and the web app with Cursor and Claude Code. The speed came from learning to run subagents in parallel: independent tasks went to separate agents (one refactoring the ingest script whilst another wrote the reporting views) instead of everything waiting on me.
Two habits kept the codebase production-ready. The repository's CLAUDE.md stayed up to date, so every agent session started with current context instead of rediscovering the project, and I trimmed the generated code we did not need.
Nothing shipped without a ground-truth comparison. Every pipeline change was checked against known outputs before it went live, so the team could trust the data without taking my word for it. I also wrote a bootstrap document that gives any coding agent the context, connections, and playbooks to run labelling tasks end to end. A small data team got a production labelling system in days.
Next.js and Supabase for the review inbox (row-level security, migration-managed schema, and reporting views), Python for the pipeline, a vision model for pre-labelling, PostHog for event tracking, and Vercel for deployment.