I build infrastructure that can explain itself: systems with explicit boundaries, reproducible results, and a record of what happened when things go wrong.
My work lives around systems engineering, agent security, developer tooling, and measurement that holds up outside a demo.
| Focus | What that means in practice |
|---|---|
| Agent systems that fail safely | Constrained tools, least-privilege credentials, controlled egress, and audit trails for consequential actions. |
| Evidence before confidence | Benchmarks include their environment and limits; unverified claims stay visibly unverified. |
| Rust at the core | Systems software, data pipelines, and tooling that are fast enough to use and simple enough to inspect. |
A local-first security runtime for AI agents. Toolcage places policy enforcement, credential brokering, controlled network access, and tamper-evident audit logs between an agent and the systems it can affect.
Rust · Linux · MCP · Security
72 tests · 404 ns policy-decision p99 · Alpha; not independently reviewed
Reproducible crypto-market measurement. Lossless market-data capture, deterministic replay, and honest cost models for research and simulation—not just backtests that look good in hindsight.
Rust · Market data · Replay
300/300 top-of-book checks · Project site
Ethereum indexing designed around failure modes. A concurrent Rust indexer with canonical-chain tracking, reorg handling, crash-safe recovery, and PostgreSQL storage.
Rust · Tokio · PostgreSQL
232 tests · 150 ns/block decode benchmark · Project site
An open evaluation harness for LLM-generated code. A reimplementation and extension of ENAMEL built to make code-efficiency evaluation easier to run, inspect, and compare.
Python · LLM evaluation · Reproducibility
13 models · 161 problems · 612 tests · Project site
- REA — making agentic reverse-engineering workflows more reliable and useful from the command line.
- ECC — contributing to an agent-harness optimisation system spanning skills, memory, security, and research-first workflows.
- DepLens — exploring whether dependency graphs, code usage, and Git history can predict what a change will break.
- Determinism before speed. Same input, same output, every replay. Optimisation only matters when the system remains explainable.
- Fail closed. If an action cannot be authorised, isolated, or recorded, it should not quietly proceed with weaker guarantees.
- Measure the boundary. Every result has a method, environment, and limit. A benchmark without context is marketing, not evidence.
- Keep the rough edges visible. Incomplete work, uncertain conclusions, and known failure modes belong in the engineering record.
