Personal project · AI engineering · In progress

Runveil

I’m building Runveil, a runtime for reliable and inspectable AI agents. My work covers the Python execution loop, PostgreSQL persistence, model integrations, controlled repository tools and an offline worker with checkpoint recovery, bounded retries and time limits. The planned application is a software-engineering agent that investigates repository tasks, makes isolated changes, runs tests and presents patches for human approval. A trace interface, reproducible evaluation and deployment are still in development or planned.

  • Python
  • FastAPI
  • PostgreSQL
  • Next.js
  • TypeScript
  • AI agent runtimes
  • Tool authorization
  • Durable execution

Agent execution

I’m building a Python runtime that coordinates model requests and tool calls within explicit permissions and execution limits. Versioned agent definitions, ordered execution history and checkpoint snapshots are stored in PostgreSQL.

Models and tools

Implemented structured model contracts, scripted and OpenAI-compatible providers, and typed tools for reading and searching explicitly allowed repository files.

Durable execution

The offline worker supports leased jobs, checkpoint recovery, bounded model retries and elapsed-time limits for controlled fixture workflows. Tests cover lifecycle transitions, persistence and recovery boundaries.

The finished experience

The goal is a software-engineering agent that takes a controlled repository task, investigates the code, makes changes in an isolated workspace, runs tests and presents a validated patch for human approval. A web console will let users inspect model and tool calls, failures, retries, latency and cost.

Evaluation and deployment plans

I plan to compare agent versions on reproducible coding tasks, reporting success rates, latency, cost and statistical uncertainty. Later phases include MCP tool integration, AWS deployment with Terraform and CI/CD, and benchmarks comparing hosted and self-hosted model inference.

Continuing development

The project is underway, with a FastAPI backend and Next.js web foundation. The finished agent workflow, approval interface, evaluation reports and public demo are planned work.