Engineering Specialization & Core Pillars
I build, evaluate, and optimize autonomous LLM agent systems with an emphasis on context engineering, local hybrid retrieval (RAG), and empirical LLM safety research. My work bridges low-level systems programming in Rust and C++ with high-throughput neural pipelines in PyTorch and transformers.
- Context Engineering & Agent Harnesses: Explicit token admission, context inspection, memory compaction, and session continuity for autonomous coding assistants.
- High-Throughput Local Retrieval: Merging Tantivy BM25 lexical indexing with LanceDB columnar vector storage via Reciprocal Rank Fusion for air-gapped search.
- LLM Safety & Red-Teaming: Controlled evaluations of prompt injection, role confusion, chain-of-thought forgery, and reasoning-channel leakage.
Selected Projects & Applied Research
- Elpis
Open-source Rust fork of OpenAI Codex CLI with explicit context pruning, memory compaction, and continuity for AI coding agents. Provides selective context admission, auditable pruning records, portable checkpoints, and provider-neutral model access.
- Eyes Wide Shut
AI safety and red-teaming research on prompt framing, dialogue state, tool use, and reasoning-channel exposure in LLMs, evaluating multi-vector jailbreak resistance on open-weight models.
- rag-mcp
High-throughput local hybrid RAG engine combining Tantivy BM25 lexical indexing, LanceDB vector search, and Cross-Encoder reranking for AI coding agents and technical workflows.
- WhisperType
Local GPU-accelerated voice input and speech-to-text with whisper.cpp CUDA transcription and optional LLM refinement for developer productivity.
- Financial Market Analysis
FX signal prediction and time-series forecasting with double-tower gated transformers, LightGBM, and CatBoost on five years of tick-level market data.