Open-sourced ML Team: analyst, engineer, skeptic agents automate the full ML pipeline
kmeanskaran · x · 2026-11-10
Developer kmeanskaran open-sourced ML Team (MIT), built with Google's ADK and Gemini, which automates an entire ML pipeline with three AI agents plus human sign-off on critical decisions.
- Roles: Analyst explores data, Engineer builds features and trains models, Skeptic hunts leaks/unfair features/weak models; the human approves features and what ships
- Agents write and run real Python at every stage instead of guessing from CSVs; the harness scores every model so no agent grades its own work
- Includes feature store, model registry with prod promotion, versioned runs, and a FastAPI /predict endpoint
- Deployable to Google Cloud with Terraform and GitHub Actions
His thesis: the future of software engineering isn't vibe-coding but automating whole teams with agents.
More from coding & agent
- Meta and Sierra Lead Working Group on AI Agent Identity Standards — mikeflache · 2026-11-10
- Stanford open-sources Meta-Harness: evolving agent scaffolds pushes Opus 4.6 to 76.4% on Terminal-Bench 2.0 — victor_explore · 2026-11-10
- Anthropic engineer earning $350K–$950K credits his edge to building agent harnesses — aftahi_ai · 2026-11-10
- Turn one animation into a repeatable pipeline: 3 aspect ratios and a Skill — FinanceYF5 · 2026-11-10
- Self-review loop: have Opus score its own frames, redo only the worst 3 clips — FinanceYF5 · 2026-11-10
- Beat-locked animation: librosa drives motion and code-synthesized SFX on one 120 BPM timeline — FinanceYF5 · 2026-11-10