Agent systems,MCP tooling,LLM infrastructure.

I built this as a Manifest V3 Chrome extension with a WebSocket gateway behind it. The agent runs on your own machine and drives the session you're already signed into, so it never needs your passwords and never has to scrape from the outside.

Lead generation is the main thing it does. Find candidates, enrich them with emails, export the result. I made the pipeline resumable with caching, deduplication and atomic checkpoints, so a run that dies at step four picks up again at step four.

It also handles receipt OCR and expense reconciliation. Credentials sit in AES-256-GCM encrypted storage. There are 204 tests covering it.

Most LLM code ends up tied to whichever provider you started with. I wanted the model to be a part you can swap out, so this runtime moves between OpenAI Codex in the cloud and Ollama or Qwen locally without the calling code changing at all.

The work was in the parts every provider does slightly differently: structured and parallel tool calls, token budgeting, context window limits. Those sit behind one interface now, and conformance checks verify each backend actually does what it claims instead of taking its word for it.

It redacts logs so they're safe to audit, and ships GPU and memory sizing utilities so you can tell whether a model fits your hardware before downloading it. 281 tests pass against it.

WalkGen is a backend MCP server built around seven tools and a fire-and-poll flow. The client starts a job and polls for the result, rather than holding a connection open for the several minutes a render takes.

Image to video runs through Kling on Replicate. A custom ffmpeg pipeline then stitches the clips together and handles transitions and frame-accurate timing. A room-aware shot grammar decides how to move through each space, because a kitchen and a hallway don't want the same camera move.

Storage is just the filesystem with a self-healing layer on top, no database anywhere. Less to run, and you can check on a job by opening a folder.