Principal Engineer, AI Systems · Incredibuild
Secure agent execution · sandboxes and snapshots · CI/build performance · computational science
I work on execution infrastructure for AI coding and research agents: isolated environments, reusable snapshots, build and dependency caches, evidence, and recovery. Earlier work spans Linux/container systems, autonomous-driving ML, and computational genomics. I hold a PhD in physics and teach software engineering at HIT.
158 merged authored PRs across 34 external repositories, 123 in 2026 (verified November 3, 2026): 151 found by GitHub search plus 7 Bazel PRs that landed through Bazel's Copybara import, which closes PRs instead of marking them merged. Full upstream catalog, all PRs and open work.
- Apache Airflow — 20 merged PRs. SandboxToolset for agent shell/file tools; Islo, Boat, OpenSandbox and OpenShell backends; CI work on compressed image stashes, snapshots and dependency cache reuse.
- OpenClaw Crabbox — 65 merged PRs. Sandbox providers, delegated-run lifecycle, lease heartbeats, identity binding and bounded parallel shards.
- OpenClaw core — 4 merged PRs. Including an SSRF mitigation for pinned dispatchers and Docker apt key-pinning fix.
- Bazel — 7 landed PRs. Merkle digest scratch buffer for remote execution, paginated, fail-closed community review labeling, locale-independent SSL monitor and a reproducible distfile.
- Also upstream: SQLFluff (13), balena-engine (11), straw (5), Omnigent (3), Zed (2), Nushell, Pydantic Monty, and more.
- Author-attributed commit integrated through another PR flow: Apache Spark UDF worker dispatch. Not included in the total above.
- ForgeFS — pre-1.0 Rust content-addressed workspaces for concurrent agents. Snapshot-pinned sessions, read-dependency checks, conflict-preserving forks, and auditable release seals. v0.4.0 released; not POSIX/FUSE or a Git replacement.
- SWFactory — alpha Airflow + agents pipeline from GitHub issues to evidence-bearing, reviewed PRs with human gates. Includes a repeatable offline fixture replay; deployed end-to-end model integration is not claimed.
- Cache Is King and Wasted Cycles — Go tooling to inspect Docker/CI cache misses, page refaults, and machine time lost during coding-agent work.
- Meta-harness on Islo — deterministic five-task harness-improvement simulation with trace visualization. Scripted agents and an Islo adapter stub; not a real-LLM benchmark.
- Gemma System One — Rust/Candle local typed-decision inference prototype with measured hardware-specific results and a documented semantic-isolation caveat.
- GULI-SET — synthetic, deterministic benchmark for distinguishing independent evidence from repeated claims; not yet an LLM evaluation result.
- Evidence-Aware MapReduce for Forkable Compute — 2026 preprint and open implementation on lineage, evidence identity and statistical reduction.
- Three-Body Orbit Atlas — numerical continuation and Floquet-stability research; in progress, with explicitly bounded completeness.
- SuperFly — connectome-informed RNN controller for simulated fruit-fly flight; research prototype.
- T2T vs GRCh38 — reproducible chromosome-level sequence-composition comparisons, not functional CpG-island annotation.
Explore all 69 active original projects and all public forks/archives.
- Airflow Summit — Graphs of AI Coding: Orchestrating Agentic MapReduce with Airflow's New Sandboxes
- Cambridge Systems Research Group — Forkable Sandboxes: The Runtime Layer for AI Software Factories
- Zenity AI Agent Security Summit, NYC — Your Agent Escaped Without Escaping the Sandbox
- LLMday NYC — The Sandbox Is the Agent Runtime: What Breaks When Execution Becomes Disposable
Computational genomics and physical biology research with Baylor College of Medicine, Rice University, and Harvard Medical School. I teach EASS Class X at HIT. See publications on Google Scholar.
Website · CV (PDF) · Writing & demos · Research · LinkedIn