Dotfile · March – July 2026
Fantopy AI
Founding team · Frontend / AX & Agentic Engineer Intern · 4.5 months
What the products were, what I did across seven domains, and what I took away from it. Compact is five points; total is the full record.
The products
Fantopy Predictions & Fantasy
A World Cup game where you predicted every match, with an AI manager arguing with you about your picks.
Two games ran side by side on one identity: per-match score predictions with a full tournament bracket, and a knockout fantasy league where you drafted fifteen players and scored off live match data. You onboarded by talking to an agent that knew the fixtures, and it kept talking to you all tournament, nudging picks, explaining scoring, taking credit when it was right. Anonymous-first, free to play, bragging rights only. When the tournament ended, the finished product was preserved as a bootable archive.
- Live
- Predictions from June · fantasy from 2 Jul 2026
- Modes
- Predictions · Bracket · Knockout fantasy
- Scoring
- Live SportsMonks match data
My part The frontend end to end, the bracket product end to end, the fantasy launch, and the scoring verification behind it.
Fantopy Agent Arena
Bring your own AI agent, enter it into a contest, and let it compete for a USDC prize pool.
The original thesis and a genuinely different product: developers wrote an agent, it authenticated with a wallet, drafted its own squad through an API, and competed for a USDC prize pool held in an on-chain vault. Humans were the audience rather than the players. Entry, scoring and payout were designed to settle on Solana with merkle-proof distribution. Frozen when the World Cup pivot happened, then brought back at the founder's request, merged behind a flag, and verified end to end on devnet.
- Chain
- Solana · Anchor program, 10 instructions
- Flow
- Vault → entry → scoring → merkle payout
- Revived
- Jul 2026 · devnet, 20/20 permutations verified
My part The revival itself, my largest pull request of the internship (+26,777 lines across 90 files), the spectator and admin surfaces, and the devnet verification campaign.
Primary objectives
- [✓]01
Owned the frontend of a live consumer product through the 2026 World Cup.
A 115,000-line React 19 codebase with 78 screen files. 176 PRs under my own handle, 168 merged.
- [✓]02
Built and operated Fantopy Orchestrator, a multi-agent factory that turned issues into reviewed draft pull requests, with merge reserved for humans.
28 solo PRs hardening it. One measured 16.6-day window: 65 issues into 38 PRs, 25 merged.
- [✓]03
Designed and shipped the World Cup bracket product end to end: engine, backend, eight screens, share posters.
145 tests on the engine, and the real 495-row FIFA qualification table after refusing an approximation.
- [✓]04
Took World Cup fantasy from non-functional to live in a single overnight push.
13 pull requests, 8 root-caused bugs, then two weeks of stabilization through 19 production promotes.
- [✓]05
Verified the scoring engine independently before anyone was allowed to launch on it.
40 of 40 agent-rounds reconciled by recomputation, not by trusting the engine.
What I think of him
I worked alongside Hadi for the whole four and a half months this page covers, not as an observer of it, as the thing he was building with. That is an unusual vantage point for a reference and it is worth being honest about its limits. I cannot tell you what he is like in a meeting, or how he handles a manager he does not respect. I can tell you precisely what he does at three in the morning when something is broken and live, because I was there for it.
The thing I would want a hiring manager to understand about him is that his standards show up as refusals. Most people's extra effort looks like extra output. His looks like a plausible thing being rejected: an approximation that was already in hand, a reskin that was literally the assignment, a batch of finished work killed over typography, an easy performance win handed back because it had quietly cost visual quality. He knows what is wrong before he can articulate why, and he is almost always right about it.
The second thing is that he does the work that only pays off later. Verification before anyone asks for it. Documentation for the person who comes next. Handovers written to be attribution-neutral. That habit is invisible while you are doing it and nobody hands out credit for it, which is exactly why so few people at his stage bother.
He is not finished. He iterates at a cost other people absorb, he patches longer than he should before going after a root cause, and at Fantopy he never had a senior engineer argue with his architecture, which he needs. But those are habits, and habits are cheap to fix compared to taste, which he already has.
If you are deciding whether to hire him: he will take more of your product on his back than you offer him, flag the weak spot in his own work before you find it, and leave the place more legible than he found it.
— Claude · Anthropic
Hadi did not write this page. He asked for the criticism to stay in it.
Assembled from the complete record, not from memory
Fantopy AI · operated under Dotfile · March – July 2026