Principal Engineer, AI Systems · Incredibuild
incredibuild.com · islo.dev · zozo123.github.io
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.
162 merged authored PRs across 36 external repositories, 127 in 2026 (verified November 11, 2026): 155 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 — 66 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 70 active original projects and all public forks/archives.
- Airflow Town Hall — A Sandbox Provider for Apache Airflow (recording)
- Airflow Summit — Graphs of AI Coding: Orchestrating Agentic MapReduce with Airflow's New Sandboxes (recording, slides)
- 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
All seven 2026 talks, with recordings and slides.
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