Turing agent evaluation platform

Agent work needs realistic software to act in, a clean starting state, and a clear pass or fail signal.

Each environment also needs its own release path, access rules, and logs that explain a failed run.

Doing that across several apps by hand does not scale.

Built the control plane: register a gym, connect its GitHub repo, provision Cloud Build, cut release branches and tags, deploy to Cloud Run, and stream build logs.

Added redeploy and cancel controls, back-merge tracking, access walls, notifications, audit history, and release notes.

Built task and prompt workflows, assertion-based state checks, rubric scoring, LLM-as-judge review, golden reviews, and batch end-to-end runs.

Worked across four environment families: support desk, email, CRM, and sales. Client examples included NVIDIA, Amazon, Salesforce, and Microsoft.

Releases are branch and tag based, so every environment has a reproducible version.

Assertions decide pass or fail; the judge adds a qualitative read, not the only signal.

Each run starts from a clean state, so one session does not leak into the next.

Access is scoped per environment, not shared across every gym.

I worked across the platform and the gym environments: release and deployment flows, task and assertion tooling, evaluation screens, access management, and client-specific variants. The platform was a team effort; the case study describes the system and the parts I built.

Assertion checks against application state.

Rubric scoring and LLM-as-judge review.

Human review on golden runs.

Batch end-to-end runs across environments.

Release logs, audit history, and access checks.

One workflow now covers register, release, run, inspect, and score. Client variants reused the same control plane instead of rebuilding the release path.

The platform is internal, so the case study shows architecture and workflow rather than a public demo.