Meta-harness on Islo
Deterministic simulation of an iterative harness-optimization loop on five toy tasks: a scripted agent, a pattern-matching proposer, and a proposed (unvalidated) Islo backend.
Selected evidence
The career story progresses through navigation, kernels, containers, physical and genomic systems, production ML, build infrastructure, and AI execution. Evidence labels distinguish public, inspectable artifacts from employer-specific production work.
PAPEROSSBENCHMARKTALKARTICLEPRODUCTIONTEACHINGPREPRINT
AI execution & build systems
Tokio CI acceleration experiment
External experiment reporting fresh compile ~46s, restored-cache ~13s, hot-cache ~3s; full test path ~75s to ~43s with restored cache. The proposal was closed as not planned; this is not adopted Tokio infrastructure.
Graphs of AI Coding
Airflow Summit 2026: agentic MapReduce, sandboxed execution, state, caching, and Airflow as an orchestration layer for autonomous software engineering.
MCP Is Not a Security Boundary
Nordic APIs Summit 2026: isolation and security boundaries for agent-to-tool execution.
Your Agent Escaped Without Escaping the Sandbox
Zenity AI Agent Security Summit, New York, October 2026: four ways a scripted agent left through doors the job opened, without breaking the sandbox wall.
Forkable Sandboxes
Cambridge Computer Laboratory Systems Research Group seminar, 15 October 2026: forkable, isolated execution for coding agents.
Evidence-Aware MapReduce for Forkable Compute
A reduction contract for snapshot-forked agent branches: workers report estimates, information, evidence IDs, and fork lineage, so one repeated error is not counted as consensus.
Technical writing
Bazel, inside and out
Three-part series on Bazel architecture, caching/remote execution, build supply-chain infrastructure, and adoption trade-offs.
GitHub Actions at large-codebase scale
Practical integration model for build acceleration beneath GitHub Actions.
The fun part was free. Proving it worked cost $1.96.
A C-to-Rust port treated as an evidence problem: disposable microVMs, deterministic checks, repeated failure, and measured compute cost.
Where Bazel hands Android kernel builds to Kbuild
Analysis of the process boundary between Bazel/Kleaf and nested Kbuild compilation.
Production ML & computational biology
Mobileye — road-junction perception
Production deep learning and global semantic understanding for road-junction perception, with emphasis on difficult and rare cases. Public evidence here covers the role and domain; internal metrics remain private.
NRGene / CRISPR-IL
AWS-published technical case study co-authored by Yossi Eliaz on a gene-editing prediction engine using EKS, SageMaker, EFS, RDS, iterative learning, and production cloud workflows.
Baylor / Rice / ENCODE
3D genome architecture, Hi-C, physical biology, graph theory, and reproducible large-scale analysis pipelines.
Harvard / HIT computational biology
Cancer transcriptomics (iScience 2026) and chemistry-informed machine learning for calcium-binding proteins (Protein Science 2025).
Teaching as an engineering artifact
EASS — Engineering of Advanced Software Solutions
HIT course spanning modern software engineering, ML engineering, cloud, containers, APIs, security, data, assessment tooling, and GTM engineering. Course artifacts include a 455-question / 34-module internal practice platform and a public automated student-project evaluator.
Systems foundations
BAE Systems / Rokar — GPS & navigation algorithms
Algorithms in a Jerusalem engineering organization publicly documented as specializing in high-end GPS receivers, GPS anti-jam, navigation, and guidance systems.
Marvell — Android/Linux kernel on PXA
Low-level Android/Linux-kernel work around Marvell PXA mobile platforms during the period in which Samsung Galaxy phones and tablets shipped on Marvell PXA SoCs.
Broadcom — firmware/software
Embedded and firmware engineering continuing the low-level systems thread.
Twistlock — containers & cloud security
A bridge from low-level systems to container isolation and cloud-native security.
Evidence policy: public papers, repositories, articles, conference listings, and benchmarks link directly to their sources. Evidence labels separate these public artifacts from employer-specific production impact.