Websparks — AI-Driven Full-Stack Software Engineer Agent

The brief

Websparks wanted an AI agent that could go beyond chat-style code completion — one that owns a feature from spec to pull request, including writing tests and updating docs.

What I built

• Multi-step planner that decomposes a feature request into discrete tasks
• Tool-using executor with sandboxed file edits, shell access, and test runs
• Long-horizon memory backed by Postgres + vector search
• Web UI for reviewing, editing, and approving the agent's plans before execution
• Telemetry pipeline for evaluating agent quality across runs

Outcome

The agent reduced manual feature-writing time significantly across tested repositories, with measurable improvement in test coverage from auto-generated specs.