Nova: an AI employee that speaks Hebrew
Outbound calls in natural Hebrew and WhatsApp replies grounded in the business's live data. Every outgoing message passes a gate that requires verification; a human approves sensitive actions.
In production
Founder of OpticoAI and Gevurah · AI Harness Engineer · Israel
Most AI agents look great in a demo and break in production. My work is the part in between: the harness.
A harness is everything around the model. It decides how work is split between agents, what each one is allowed to do, what it remembers, how its output is checked, and when a human has to approve. Get that right and agents can safely touch real customers, real money and real inventory.
Not demos. Every system below runs today against real customers, money or inventory.
Outbound calls in natural Hebrew and WhatsApp replies grounded in the business's live data. Every outgoing message passes a gate that requires verification; a human approves sensitive actions.
In productionWeb chat, WhatsApp and Messenger. Product and order questions, quotes for kindergartens and schools, abandoned carts and shipping updates, all from live Shopify data.
In productionRun daily against supplier and competitor prices and change live store prices. Canary before every full run, additive changes only, non-regression check before anything ships.
In productionClaude plans, Codex executes, Gemini reviews. A task contract for every job, a signed trust level per model, human approval per exact action, and ~4,700 work orders processed so far.
In productionProduct research, copy, images, Hebrew voice-over, video, and actual publishing to Facebook, Instagram and YouTube, including article SEO and product-page embeds.
In productionA receipt photo sent on Telegram becomes an indexed invoice, a ledger entry and a report ready for the accountant. Supplier billing, receipts and income in one place.
In productionColoring pages, a live holiday calendar and materials for kindergarten and school teachers, with a live download counter and full measurement. A free service that doubles as an SEO engine.
In productionAn AI studio for ecommerce marketing content. Built fast, then moved onto infrastructure I own: Cloudflare, Supabase, rehearsed backup and restore.
Closed pilotA harness that measures whether a coding agent succeeds every time, not just once. How I test models before they touch a real system. github.com/oreno334/passk
Open sourceEverything above runs on one system built from scratch since May 2026. Measured directly from the repository on 31 Oct 2026, git-tracked code only.
Most of this code was written by models, under my direction. That is the point: the skill is architecture, task decomposition, writing specs a model can execute, and building the verification that catches the times it was wrong.When an autonomous system touches customers' money, the interesting question is not what it can do, but what you stopped it from doing.
15+ years in ecommerce, from running digital at Office Depot Israel to leading growth for online retailers. I founded OpticoAI and Gevurah, and my own Shopify store, Shopping Zol, is the live lab where every system is tested on my own business first. Open-source: passk, a harness that measures whether a coding agent succeeds every time.
The full profile, portfolio and numbers are on the Hebrew page.