Services
Three lanes. Every engagement starts with a written plan before any code is written.
Python pipelines and tooling
Replay pipelines, data plumbing, experiment harnesses, and workflow tooling that stays reproducible and usable after handoff.
Output: runnable baseline suite, scripts, and a results workflow you can extend.
LLM integration and agentic systems
RAG pipelines, MCP integrations, long-context system design, and tool-augmented agents — built as proper engineering artifacts with evaluation harnesses and explicit failure modes.
Output: working system, evaluation harness, and a written analysis of trade-offs and risks.
Systems programming in C++ and Rust
Design and delivery of fast, reliable systems from the ground up. Profiling, hot-path redesign, lock contention, allocator strategy, and network tuning where needed.
Output: production-ready code with tests, benchmarks, and a written explanation of design decisions.
What I focus on
Performance-critical problems where the cost of getting it wrong is high.
Projects
Selected public work. Additional client work and private projects available under NDA.
EEGL
Explanation-Enhanced Graph Learning
Research framework that improves GNN accuracy through an iterative feedback loop: run a GNN, extract structural patterns from its explanations, inject those patterns as new node features, and repeat. Makes GNN explanations actionable rather than merely interpretive.
Python · PyTorch · PyTorch Geometric · Gaston subgraph miner · Glasgow Subgraph Solver
Published: arXiv 2403.07849
Substrate
Open-source OMS/EMS simulation platform
C++ limit order book and matching engine using SBE wire protocol (the same binary encoding as CME iLink), with a Python layer for fill simulation, historical replay against real Databento tick data, and DAG-based workflow orchestration. Built for reproducible execution research.
C++ · CMake · SBE · Python · Polars · Databento
Domain-specific RAG system
Confidential client project
High-confidence retrieval-augmented generation pipeline for a specialized domain where retrieval accuracy is business-critical. Built with explicit evaluation harnesses, per-query confidence scoring, and failure classification. Designed to fail loudly rather than hallucinate quietly.
RAG · MCP · long-context design · evaluation harness
Contact
Send a short note with context. I respond faster when the ask is crisp.
Opens your email client.
Include: timeline, stack, constraints, and what “done” looks like.
Availability
- Contract: fixed-scope or retainer
- Full-time: systems or execution research
- Location: Chicago-based · onsite in New York and London · remote
- Next available start: March 2026
If you need an NDA first, say so.