AI Deployment Readiness API

One-Line Summary

Public-safe FastAPI mini-project that demonstrates how I think about AI deployment readiness: evidence ingestion, retrieval, evals, release gates, rollback plans, Docker, and CI.

Why I Built It

TeachClaw and OpenClaw contain the strongest real evidence, but not all of that code or operational context should be public. This project creates a small, clean artifact I can point hiring teams to without exposing private product code, teacher data, runtime secrets, or internal OpenClaw setup.

The goal is not to pretend this is an enterprise platform. The goal is to show the shape of production AI deployment work in a repo a recruiter or engineer can understand quickly.

What It Proves

• Python/FastAPI service design with explicit request/response schemas.
• Document ingestion and retrieval with stable source IDs.
• Eval runs that can require expected terms and expected source documents.
• Deployment registration with owner, environment, rollback plan, risk notes, and pass-rate threshold.
• Readiness endpoint that returns pass, needs_judgement, or block.
• Health and metrics surfaces for operational inspection.
• SQLite-backed persistence for documents, deployment profiles, eval runs, and audit events.
• Optional API-key protection for write endpoints.
• Audit endpoint and structured JSON write logs.
• Dockerfile, docker-compose, Makefile, and GitHub Actions CI.

Why It Maps To AI Deployment Roles

AI deployment roles rarely fail because someone cannot call a model API. They fail when teams cannot tell whether a system is ready to launch, which evidence is current, which evals matter, what rollback means, and whether the live system is using the version they think it is using.

This project turns those instincts into a small public system:

1. Register the deployment.
2. Ingest the evidence or documents the system relies on.
3. Run evals against the behavior that matters.
4. Record an audit trail of deployment/eval activity.
5. Check readiness before launch.
6. Inspect health, metrics, and recent audit events.

Stack

• Python
• FastAPI
• Pydantic
• SQLite
• Pytest
• Docker
• GitHub Actions

Honest Scope

Current implementation uses deterministic token-overlap retrieval and SQLite state. That is intentional for a compact public proof. The next real production step would be hosted deployment, embeddings/vector search, rate limits, stronger service auth, and immutable eval reports.

Repo Path

mini-project/