Projects
Systems I've built, mostly where model behaviour meets a hard constraint — a voice agent held to a latency budget, retrieval that has to stay honest over real documents, a gateway that has to fail over without the caller noticing. Linked titles have a write-up.
An agent that runs the first pass of an incident: when a CloudWatch alarm fires it checks logs, recent deploys, metrics, and runbooks, then posts a root-cause summary to Slack for one-click approval.
An interactive Recursive Language Model playground that visualizes how an LLM explores large documents, delegates subproblems to recursive calls, and synthesizes results through a step-by-step execution tree and timeline.
A macOS desktop app that transcribes live calls, detects customer objections, suggests follow-up questions, generates meeting summaries, and coaches the rep while the call is still running.
A terminal coding agent that streams a model, calls file tools, and runs every tool inside a locked-down Docker sandbox. The same loop works with OpenAI and with Gemma through Ollama.
A RAG system that lets users upload PDFs and get accurate, context-grounded answers through semantic search and LLM question answering.
A single API interface across multiple LLM providers, with request routing, model selection, API key management, and usage tracking.