Full-stack & ML engineer. I build production systems end-to-end — backend, frontend, and the infrastructure under them — and I work AI-native: LLM agents, agent memory systems, MCP tooling, local models.
| repo | what it is | stack | status |
|---|---|---|---|
| procurement-forecasting | demand forecasting built around the ordering decision — quantile-of-sum orders, tiered service levels, order-level backtesting; benchmarked on M5 | Python · Nixtla · LightGBM |
v1 |
| tenebra | cross-platform VPN client on sing-box — stdlib-only core, honest leak-check, CI with race detector + e2e; ships for Windows, macOS and Linux | Go · Tauri · React |
v0.4.6 |
| claude-memory-template | file-based persistent memory for CLI coding agents — layered markdown vault, always-on index under a hard 17KB budget, CI-enforced entropy control | Markdown · Python · git |
v1 |
| Dota AI Coach | real-time coaching over Valve's official GSI — deterministic rule engine + LLM advice on a transparent overlay | Python · FastAPI · Electron |
early |
- applied ML — demand forecasting on the Nixtla stack: demand classification, conformal prediction intervals, hierarchical reconciliation, honest leak-free backtesting
- LLM tooling — real-time agents, MCP integrations, local-model pipelines, and persistent memory that keeps agents useful across months of sessions
