personal rag — a control room for task-specific personal-ai instances
chromadb over my own claude exports, browsing history, identity interviews, and a small books library (~14k chunks). served as an mcp tool so other agents can query it. designed as an observer sitting above task-specific instances (journalling, voice, gateway) that report back into the same index.
slovak driving-license theory exam study app
desktop study app over a 951-question canonical bank scraped from minv.sk. drill mode, leitner-style spaced repetition, 40-question exam simulator with timer, browse + difficulty labelling, 52-week heatmap stats. ~270 tests, 93% coverage.
private reflective-conversation partner with graph-rag memory
local-only chat ui that journals into its own private corpus. retrieves verbatim chunks (dense / bm25-hybrid / cross-encoder reranked) and a sqlite entity-relation graph distilled per conversation. 3d umap + hdbscan cluster panel and a force-directed graph panel for self-inspection. crisis filter + safety rails.
historical-market sim & trading-bot tournament
simulator + roster of trading bots (math baselines → ml → rl) competing on historical ohlcv with a configurable cut-off, fixed budget, and a leaderboard ranked by sharpe / max-drawdown / cumulative return. walk-forward cross-validation hyperparameter search runs as slurm arrays on the dais b200 cluster.
curated en-wikipedia corpus built on hpc
world-knowledge complement to the digital avatar. yaml-defined category bundles → category crawl → article fetch (slurm array) → filter + chunk → bge-m3 embed on b200 arrays → chromadb. served either locally (small bundles) or as a remote mcp behind an ssh tunnel (large bundles).