For Zach Yadegari · the Persona CTO role

Hi Zach.
I want to be your CTO.

I make AI agents reliable and cheap. Persona is free, so every task it does is your cost. I get cheap models doing big-model work.

My Persona test: a pizza order in Polish, right the first time. Five hotel calls took 40 minutes, and I only heard back when I asked. That’s what I’d fix first.

Talk to it, it’s live. The other two show how I’d build them.

9:41
Persona
tap a task
Talk to it, or tap one of the others. This panel shows what happens underneath.
You asked for three

What I’m proudest of.

loading the run…
Level 1 / 9
0 / 734 moves
0 model picks
0 restarts
level cleared forecast missed (4 of 417)
  1. An agent that learns games with no instructions

    ARC‑AGI‑3: puzzle games with no rules given. My first version was tuned on the public games and scored 0 on new ones, so now test games stay held out from day one. The rebuild has no game‑specific code: with GPT‑6 Astra it cleared all 183 public levels, and a cheap open model cleared whole games cold, this one for 6.5¢. My code builds the options, the model picks, and a checker replays every move.

  2. Models that keep what they learn

    A small model moves new facts into its own weights overnight and keeps 88–93% with its notes wiped. Plain fine‑tuning stays at chance.

  3. Trendvo’s whole core

    Co‑founder, leading tech. I built the fintech end to end: broker connections, order flow, scoring. I’d never built a trading platform. I started and figured it out.

Also: a solo paper on arXiv, SpecWise (CI checks for AI‑written code) and the ARC runs, move by move.

Week one

One thing a day.

  1. Day 1
    All 5 hotel calls in the time of one
    Five at once with a live transcript, so the total is the longest call, not the sum. My test took 40 minutes.
  2. Day 2
    Mid-call questions come to your texts
    The hotel stays on the line while you answer. On the band once it ships.
  3. Day 3
    Phone menus, learned once
    Gets through a menu once, saves the path, never waits on it again.
  4. Day 4
    Every step on the smallest model that passes
    Cost and time-to-first-word per task on one chart, both going down.
  5. Day 5
    Our own model, first run
    A ~3B open model trained on opted-in task traces, names and numbers stripped. It takes the easy steps, so fewer go to OpenAI or Anthropic.

Worth 20 minutes? I’ll screen-share the ARC harness, go through my notes on Persona, and show how I’d get the 5 hotels down to the time of one call.

Pick a time

Questions

The short ones are here. For anything else, ask below.

Next to Tanay, full-time. He knows Persona’s calls and stack better than anyone; I’d take whatever he wants off his plate, like servers, on-call, shipping and hiring, so he can keep building.

Yes, Python and TypeScript. I designed the ARC harness: the part that builds the options and the checker that replays every move with zero model calls. Coding agents write a lot of the code to my specs and tests; I read it, debug it and own it. I’ll screen-share any file on a call.

Yes, whenever you need. I can start the day you say.

Most of what I build is private: the fintech’s code belongs to the company, and my research loop isn’t open-sourced yet. What’s public is my paper’s code and this site. Happy to screen-share any of it.

A cheap open model cleared 4 ARC-AGI-3 games cold, bp35 at 9 of 9 for about 6.5¢. With GPT-6 Astra my loop cleared all 183 public levels, all 25 games, every move replayed.

20, Polish, based in Warsaw. I was at VU Amsterdam until July 2026, when I left to build full-time. I’m co-founder leading tech at Trendvo, a Sydney fintech, and I wrote a solo arXiv paper (2604.03266).