Projects

Projects by Masih Moafi spanning AI coding agents, context engineering, LLM safety and red-teaming, retrieval, developer tools, and applied machine learning.

  • Elpis

    A Codex-based coding environment with explicit control over an agent’s working state.

    • Context: inspect inputs; prune tool output.
    • Continuity: resume from saved checkpoints.
    • Work graphs: order tasks by dependency.
  • Eyes Wide Shut

    Three controlled gpt-oss:20b safety studies with published prompts, outputs, and limitations.

    • Tool calls: 78/100 → 97/100 with simulation.
    • Reframing: education label bypassed refusal.
    • Secret leakage: 25/30 reasoning traces.
  • rag-mcp

    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.
  • Financial Market Analysis

    A EUR/USD signal-research pipeline, from five years of minute bars to MetaTrader 5 backtests.

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

    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.
  • WhisperType

    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.
  • After Three Years: Rebuilding a Disaster Tweet Classifier

    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.