# Yossi Eliaz: Writing & experiments

> Yossi Eliaz: Engineering & Growth, Incredibuild / islo.dev. Current work: AI execution infrastructure (agent sandboxes, CI runners, caching, build acceleration) as Principal Engineer, AI Systems at Incredibuild / islo.dev. Earlier work: computational genomics (3D genome, Hi-C, ENCODE pipelines, CRISPR), autonomous-driving perception at Mobileye, and physics (PhD, University of Houston). Senior Lecturer at HIT.

This site collects Yossi Eliaz's writing, field guides, and live experiments. Professional profile, CV, and publications: https://yossieliaz.netlify.app/ (research index: https://yossieliaz.netlify.app/research/).

## Selected work

- [ForgeFS](https://github.com/zozo123/forgefs): Isolated, auditable workspaces for concurrent coding agents. When two agents publish over each other, the second one's work survives as a fork.
- [SWFactory](https://zozo123.github.io/ariflow-swfactory/): GitHub issues to reviewed pull requests with coding agents, Airflow, and human approval gates. The local replay needs no keys.
- [What T2T changed](https://zozo123.github.io/t2t-vs-grch38/): GRCh38 against the telomere-to-telomere T2T-CHM13 reference, chromosome by chromosome: +179.6 Mb of usable sequence and where CpG-rule windows fall. Composition only; caveats published.

## Featured essay

- [Databases in the Age of AI](https://zozo123.github.io/databases-in-the-ai-era/): What fifty years of database systems must become when the dominant reader of data is no longer human.

## AI Agents & Sandboxes

- [The Sandbox Shift](https://zozo123.github.io/sandboxes-why-how-when/): When isolated execution helps agent systems, and how to choose an approach, with an interactive decider.
- [CI Caching Is Not One Cache](https://zozo123.github.io/gha-cache-field-guide/): A June 2026 POC: when native caches, BuildKit, or Incredibuild Build Cache help GitHub Actions on fresh, warm, or touched runners, with POC timings (not a controlled benchmark).
- [Wasted Cycles](https://zozo123.github.io/wasted-cycles/): Find the machine time blocking agent work: local traces and GitHub Actions latency, with a one-line installer.
- [Hermetic Agents](https://zozo123.github.io/hermetic-agents-demo/): Deterministic demo: scripted code and test writers get the same spec but disjoint inputs; a QA arbiter with a hidden oracle blames the seeded wrong test. No language model is called.
- [Bazel 9 × monty](https://zozo123.github.io/bazel-monty/): Building pydantic/monty with Bazel 9 and a remote cache: 121 s cold, 15.7 s fully cached.
- [AI Sandbox Landscape](https://zozo123.github.io/sandbox-landscape/): A map of the execution-sandbox ecosystem.
- [One Lab Per Kernel](https://zozo123.github.io/sol-exec-bench/): A technical design for running a GPU-kernel research loop (SOL-ExecBench) reproducibly on Islo sandboxes.
- [Rust It Up](https://zozo123.github.io/rust-it-up/): Prototype: paste a public GitHub repo to get an illustrative Rust-migration scorecard. The scanner is mocked; estimates are not audited.
- [Inferoa Flight Recorder (unofficial demo)](https://zozo123.github.io/inferoa-demo/): Unofficial demo of agentic-in/inferoa: a client-side simulation of prefix caching, context compression, and routing, plus receipts from real vLLM runs in islo.dev sandboxes.
- [Agentic Airbyte](https://zozo123.github.io/agentic-airbyte/): Design for agent-driven ETL, plus one verified run: 50,000 rows from ClickHouse to DuckDB in an islo.dev sandbox. Not affiliated with Airbyte.
- [Evals or You're Blind](https://zozo123.github.io/demo-evals-pydantic-islo-crabbox/): Runnable demo after Joshua Görner's PyCon DE 2026 talk: Pydantic Evals grade a Pydantic AI build-triage agent offline, live, or on a disposable crabbox box.
- [Repo2RLEnv × Crabbox](https://zozo123.github.io/repo2rlenv-crabbox-demo/): Demo record: scoring Repo2RLEnv pr_diff tasks on an islo.dev sandbox via crabbox, about 50 s per task.
- [Odysseus × Crabbox](https://zozo123.github.io/odysseus-crabbox-demo/): Demo: run odysseus-dev's Odysseus AI workspace, and its agent's code, on throwaway islo.dev microVMs via openclaw/crabbox, from a fork's one-file crabbox.sh.
- [Tiny Long-Memory Benchmark (toy demo)](https://zozo123.github.io/longmem-mini-on-islo/): A LongMemEval-style toy benchmark fanned out across sandboxes.
- [Meta-harness on Islo](https://zozo123.github.io/meta-harness-on-islo-page/): A deterministic simulation of harness optimization: scripted agent, pattern-matching proposer, five toy tasks.
- [env-rosetta](https://zozo123.github.io/env-rosetta-page/): Demo: @adithya-s-k's Wordle env from RL_Envs_101 in four RL frameworks (OpenEnv, ORS, NeMo Gym, Verifiers); two ran in islo.dev sandboxes, and the sample rollouts are scripted.
- [Crabbox Runner for RustRover](https://zozo123.github.io/rust-rover-plugin-crabbox/): MVP plugin: run cargo test and clippy from RustRover through the OpenClaw Crabbox CLI on remote Crabbox runners or Islo sandboxes.
- [Gemma System One](https://zozo123.github.io/gemma-to-jev/): Research prototype: Gemma 3 4B as a one-forward-pass decision function in native Rust, with shared-state inference across many questions.
- [Pokeloop](https://zozo123.github.io/pokeloop/): Demo: a genetic-algorithm loop over LLM system prompts, designed to fan out across parallel islo.dev sandboxes, with a Pokémon Crystal-style viewer. Results are a scripted playback; no agent was trained.
- [unity-loop](https://zozo123.github.io/unity-loop-page/): Proof of concept: Claude proposes and scores CSS variants around one fixed Unity WebGL scene (Lightning VFX by @MirzaBeig), each served from its own islo.dev sandbox. Scores shown come from a deterministic offline demo.
- [wolfram-fb0](https://zozo123.github.io/wolfram-fb0/): Can a 1B local model write x86-64 assembly for Rule 30 inside a VM? First 3 iterations: not yet.

## Science & Genomics

- [Reference Broadcast by VM Snapshot](https://zozo123.github.io/genomics-sandboxes/): Demo on islo.dev: warm one sandbox with the GRCh38 reference, snapshot it, then fork it copy-on-write per chromosome for genomic scatter-gather.
- [Open Physics Problems](https://zozo123.github.io/open-physics-problems/): 112 open physics problems mapped like open mathematics: publication cultures, refutability, and ten-day agentic sprints.
- [Evidence-Aware MapReduce](https://zozo123.github.io/boltzmann-mapreduce/): Fork workers freely while keeping evidence, uncertainty, and shared origins visible, so a repeated error is not counted as consensus. Preprint, 2026.

## Deep Dives & Essays

- [The Stochastic CPU](https://zozo123.github.io/stochastic-cpu/): A field report on the cloud being rebuilt for agents: the agent hour, two-layer caches, snapshot provenance.
- [From Microservices to Microcells](https://zozo123.github.io/microservices-to-microcells/): A plain explanation of microcells: tasks that run when needed, save progress, and survive a crash, plus a safe migration path.
- [Free* Intelligence](https://zozo123.github.io/free-intelligence/): A field report on the difference between a fluent model answer and an accountable agent harness.
- [The Living Layer](https://zozo123.github.io/the-living-layer/): A speculative architecture thesis: sandboxed execution plus governed, queryable state.
- [The Data That Taught the Machines](https://zozo123.github.io/how-stackoverflow-taught-ai/): How Stack Overflow trained its own replacement.
- [The Loop Is Coming Apart](https://zozo123.github.io/loop-coming-apart/): jj, Crabbox, and the unbundled development cycle: the diff is the unit of work.
- [Intel: Vision Without Execution](https://zozo123.github.io/intel-story/): An opinion essay in comic style on Intel's bets and misses, 2000–2026.

## Talks

- [Your Agent Escaped Without Escaping the Sandbox](https://github.com/zozo123/nyc-talk): Four doors an agent can walk out through without breaking the wall: the key, the folder, the wire, the verdict. Zenity AI Agent Security Summit, New York, October 2026.
- [Forkable Sandboxes: The Runtime Layer for AI Software Factories](https://github.com/zozo123/cam-talk-london-26): Seminar at the Cambridge Computer Laboratory Systems Research Group, 15 October 2026: forkable, isolated execution for coding agents. Slides, scripts and references.
- [A Sandbox Provider for Apache Airflow](https://zozo123.github.io/airflow-sandbox-townhall/): One ephemeral microVM per Airflow task try, for fast fan-out and controlled untrusted execution. Airflow Town Hall, July 2026.
- [A Million Compiles. One Robot Hour.](https://zozo123.github.io/rust-china-conf/en/): Rust China Conf talk page in English (中文 version linked): a simulated robot, a stale-observation bug, the Rust fix, and the Incredibuild rebuild loop, with recorded evidence.

## Quant Finance

- [How Market Making Works](https://zozo123.github.io/market-making-sandbox/): 13 in-browser simulations, from the limit order book to Avellaneda–Stoikov, Glosten–Milgrom, and Kyle's λ.
- [Avellaneda–Stoikov, in a Sandbox](https://zozo123.github.io/microprice-sandbox/): A synthetic reproduction of the Avellaneda–Stoikov vs symmetric market-maker comparison from Sasson, Ho & Samson (Stanford MSE448, 2018), run in an Islo sandbox in about 90 seconds with no local Python. The qualitative result reproduces; the dollar magnitudes do not.

## Profiles

- [GitHub](https://github.com/zozo123)
- [LinkedIn](https://www.linkedin.com/in/yossi-eliaz/)
- [Google Scholar](https://scholar.google.com/citations?user=NL1ZyOgAAAAJ)
- [X](https://x.com/YossiEliaz)
- [Professional site](https://yossieliaz.netlify.app/): profile, CV, and publications