I build AI systems and the products that carry them — an autonomous agent that runs the same way on every major model, and Hebrew right-to-left software shipped end to end: storefronts, editors, a PDF workspace, a lead pipeline that signs what it stores.
Independent, in Israel. Everything below is live, and every number on this page is generated from a run rather than typed.
Most of my AI work lives in fabius — one autonomous agent, fifteen zero-overlap skills, running the same way on Anthropic, OpenAI, Google, Mistral and Groq. The agent is the easy half. The apparatus around it is the work.
|
Every benchmark task is answered three ways — the bare model, a "be concise" control, and the shipped stance verbatim — then scored by two blind judges that are never told the model or the arm. Across four Claude tiers the output cut is 20–35% on every model, and on every capable tier the agent beats the bare model and the control. The one regression is printed in the table rather than dropped: it is the empirical case for routing by model tier. |
Every release is sealed: a SHA-256 Merkle root over each skill contract, an Ed25519-signed git tag, and an OpenTimestamps proof anchoring the commit into Bitcoin.
|
Behaviour is a file, not a prompt — a routing core, an identity contract, and fifteen skill contracts the agent loads verbatim. The same frontmatter contract is read by Claude Code, Codex and grok-build, so one folder is discoverable across three harnesses. A zero-dependency Node runtime carries the same rules offline under 75 tests. |
|
AI specification and implementation for businesses My agency. Twenty-three Hebrew RTL pages that sell the actual work — site builds, CRM, automations, agents and GEO — with matched depth on every service page, a Gemini chat agent that captures a lead inside the conversation, and an HMAC-signed lead pipeline on Cloudflare Workers, D1 and Resend.
|
The autonomous agent that runs on every major model Fifteen coordinated, zero-overlap skills under one stance — scout wide, strike narrow. A 22-rule decision core, an IDENTITY contract and the FBS benchmark, running the same way on Anthropic, OpenAI, Google, Mistral and Groq. Version 2.3.1, cryptographically sealed, and installable by anyone:
|
|
A Hebrew PDF workspace that never uploads your file Sign, stamp, merge, compress and edit — the whole flow runs in the browser, so no document ever reaches a server. Carries a verifiable SEAL that binds a file to a hash anyone can re-check, behind Turnstile and a Cloudflare Worker edge.
|
Browser-only visual HTML editor Design a landing page visually and export clean HTML — nothing to install. Seventy brand templates and an AI builder that writes in 142 distinct design-system voices, while a zero-styling guardrail keeps your layout intact. RTL-aware and fully client-side.
|
|
42 privacy-first developer tools Hash, encrypt, sign, derive, generate and convert — from HMAC and Argon2id to Shamir secret sharing, PASETO and post-quantum ML-KEM/ML-DSA. Every byte is processed in your own browser: nothing is uploaded, nothing is logged, and it keeps working offline.
|
Eleven case studies, each with a live demo A no-build Hebrew RTL site on a light Liquid Glass canvas. Every case study opens both a written study and the working product it describes, alongside long-form writing on AI and agents, an HMAC-gated Cloudflare lead pipeline and a matching light-glass AI assistant.
|
Behaviour is a file, not a prompt. The agent's stance, its routing rules and its fifteen skill contracts are versioned documents it loads verbatim. A prompt you cannot diff is a prompt you cannot fix.
Nothing ships on "looks right". A change is checked against the running system, not the source — headless browser, real URL, console and network asserted. Code reading is fine for fixing and not for concluding.
Claims carry receipts, including the ones that hurt. The benchmark is blind-judged with a control arm, and it prints the one model that regressed. Releases are sealed with a Merkle root, a signed tag and a Bitcoin-anchored timestamp. Three versions of the provenance paper were wrong, and the corrections are in the paper rather than behind it.
Every artifact is written twice. Once by the author, once by a reviewer whose only job is to disprove it. On the last pass over these repositories that reviewer pulled 63 unsupported claims — including some the source notes themselves had wrong.
A trustless anchor for document provenance ↗
A signature proves who. It does not prove when — and in the generative-AI era, when is the field that decides whether a document is evidence or a fabrication. A 72-page position and architecture paper on sealing files so their provenance survives adversaries, transcoding, hash breaks and decades of time.
It makes exactly three claims and is explicit about what it does not claim: the hard/soft binding of what, who and when belongs to C2PA and JPEG Trust, and an earlier draft claimed it as novel before review established it was already an ISO standard. The published reference package regenerates every figure in the paper — 15/15 in-EVM assertions on a registry contract with no owner and no upgrade path, a 17-vector conformance suite with real ML-DSA-65, and the benchmarks — so nothing has to be taken on faith.
Solidity EIP-712 OpenTimestamps Certificate Transparency ML-DSA-65
Every image on this page — the live stat bar, the project cards and the snake — is generated and served from this repository. No third-party README services.