Local-first personal AI memory layer that syncs context across ChatGPT, Claude, Poke, and Claude Code. A 4-agent pipeline extracts durable facts from conversation exports with Claude structured output, detects conflicts, and propagates approved memories back to each platform in its native format — backed by SQLite, sync-safe markdown, and a decaying workspace served over MCP.
Trajectory-aware task recommender using a 4-stage LLM pipeline inspired by Self-Guided Self-Play, with embeddings-based learning (text-embedding-3-small), logistic regression re-ranking on 1545-dim feature vectors, and an evaluation framework that scores predictions against actual commit diffs.
Voice-agent assessment platform for staffing and recruiting teams. Recruiters author role-specific phone screens with AI one-shot generation and natural-language edits, candidates complete token-gated voice interviews with workspaces, and the backend turns transcripts, workspace history, behavioral signals, benchmarks, and anti-gaming checks into queryable evaluations.
Simulation-driven poker strategy lab with Random Forest EV regression, supervised feature vectors, natural-language strategy compilation, and leak detection for custom agents.