Projects

A Rust coding-agent environment built on Codex, with terminal tools, code editing, and resumable sessions.

• Code changes: inspect files and apply patches.
• Terminal tools: sandbox and approval controls.
• Session continuity: resume a coding conversation.

LLM safety research on framing, reasoning leakage, and adversarial tool-result ordering.

• Eyes Wide Shut: framing, tool use, and reasoning leakage.
• ChronosAttack: authentic tool results, adversarial ordering.
• Evaluation: controlled comparisons, defenses, and ablations.

An MCP server that finds relevant passages in local files and code without loading the whole workspace.

• Hybrid search: keywords plus embeddings.
• Ranking: fuse results, then rerank.
• Code corpus: full answers for 30/33 queries.

EUR/USD research comparing a dual-tower Transformer with tree models in backtests.

• Labels: profit, loss, and time barriers.
• Models: dual-tower GTN versus boosted trees.
• Execution: next-bar entries in MT5.

Versioned procedures that let coding agents reuse your working methods instead of relearning them.

• Triggers: load the relevant workflow.
• Guardrails: explicit steps and checks.
• Portable files: inspectable SKILL.md text.

Local dictation for Linux/NVIDIA: speak, transcribe, and paste directly into the active app.

• Capture: start and stop with a hotkey.
• Transcription: Whisper or Parakeet on GPU.
• Cleanup: optional Gemini text refinement.

Rebuilt a tweet classifier to measure how each modeling choice changes validation performance.

• Baseline: verified by overfitting one batch.
• Custom Transformer: 78.07% validation.
• DistilBERT: 83.32% peak validation.

Obsidian but fast. A native Rust Markdown editor with local vaults, linked notes and graphs. Early Linux alpha.

• Native Rust: familiar Markdown editing.
• Local files: portable vaults and note links.
• Open source: contributors welcome; parity work continues.