Backend engineer to build a CI regression service for AI coding agents (.NET / Python / Azure / AWS)

**Summary** I am looking into building a backend service that runs a team's AI coding agent against their own private repository in isolated sandboxes, grades each run with hidden tests, and reports whether a prompt or model change made the agent better or worse. It is API-first: a REST API with webhooks, a GitHub Action, and a CLI. The work includes several components: the control-plane API and job runner, the GitHub App and check, run comparison and scheduling, and hardening and deployment with Terraform. **Example of the finished product in use** In one sentence: CI for an AI coding agent. The team has a test suite for their code; we give them one for the agent that writes their code. We do not build the agent or the model; we run someone else's agent in isolation and measure it. A team ships a bot that takes a GitHub issue and opens a pull request. With our service: 1. They register their repository, their agent (a Docker image and a command, using their own model key), and a set of tasks: past issues, each with a hidden test. 2. A developer changes the agent's prompt and opens a pull request. 3. We run the agent on every task in its own sandbox and run each hidden test. 4. The PR gets a check: pass rate against main, which tasks regressed, which were fixed, and a link to a JSON diff. 5. The developer reads the diff, fixes the prompt, and pushes. **Constraints** - Language: C# (.NET, ASP.NET Core), Python (FastAPI), or TypeScript (Node). - Cloud: Azure or AWS. Managed containers, a queue, Postgres, object storage, and the cloud's secret store (Container Apps and Service Bus, or Fargate and SQS, with their equivalents). Terraform either way. - Database: PostgreSQL for suites, runs, and results. Object storage for trajectories and artifacts. - Frontend is minimal. The GitHub check and the JSON are the interface in the first version. **What exists today** Architecture notes, the API resource list, and a task format. No code.