GOAT.AI2026 — present
GOAT.AI is a platform that turns developers’ sessions with AI coding agents into structured datasets for AI labs. It captures work in Claude Code, Codex, Cursor and Antigravity, preserving the sequence of steps an agent takes while working on a coding task.
The platform handles the full path from session capture to dataset preparation: collecting new sessions and importing history, encrypting raw traces, redacting personal information and removing duplicates. The resulting datasets contain agent trajectories from real development work.
Founding Engineer and the sole engineer at GOAT.AI. I designed, built and shipped the platform in under four months, from an empty repository to production. I continue to develop and operate it.
Architecture & operations
- Development
- 600+ commits. A TypeScript, Python and Go monorepo with 9 apps and 24 packages. Built AI-natively with parallel Claude Code agents.
- Ingestion
- OpenTelemetry via Fastify HTTP and gRPC. BullMQ workers, with PostgreSQL, ClickHouse and Cloudflare R2 for storage.
- Privacy
- Per-project XChaCha20-Poly1305 encryption of raw traces in an isolated database. A default-deny PII-redaction service built with Presidio and spaCy.
- Production
- Hardened for bursty load with load shedding, admission control and separate live and backfill queues. Migrated infrastructure with zero downtime.
