Seventeen, in my final year at Letovo School in Moscow. I build forecasting models for UFC and boxing and use them to test how efficient market prices are. I’m also a co-founder and Head of AI at Gluline, a messenger.
One pre-registered hypothesis needs an e-value of 20. Test k at once and the first needs 20 × k: for the 84 in my first paper, 1,680. Drag the cursor.
Two forecasting systems, a messenger, a trading pipeline for a client, two open-source tools and a bank of 2035.
Live product at vertexmma.com
A nine-model UFC forecasting stack, measured against the market’s closing price. It covers 4,581 fighters, 8,906 bouts and 795 events.
0.7244
| Measure | Model | Closing price |
|---|---|---|
| Accuracy higher is better | 0.6753 | 0.6838 |
| Log-loss lower is better | 0.6171 | 0.5922 |
| Brier lower is better | 0.2140 | 0.2035 |
The price is ahead on all three. The model is better calibrated than the price and less sharp.
Released 19 August 2026. Co-founder, engineer and Head of AI.
A messenger that says exactly who can read what, on the App Store, Google Play and the web, in ten languages. From the first files in April 2026 to the App Store took four and a half months, with 309 unit tests at the release gate.
| Chats | Encrypted on the device, with a key for each chat |
|---|---|
| The AI assistant | Off until you switch it on; a chat’s own settings can keep it out of that chat |
| One-to-one calls | Encrypted between the two devices |
| Group calls | Not end-to-end encrypted |
Concluded research, September 2026. The report and the code are public.
It began from a thesis:
Regional and club boxing is priced less efficiently than the top of the sport.
On every bout with a market price, the data says the reverse: the price moves furthest towards the model at the top of the sport, and least on club cards.
Five forecasts of the same 3,077 fights
Log-loss from June 2023 to December 2025; further right is the better forecast. The opening price is posted a median three days before the fight.
How far the price moved towards the model, by the level of the fight
Bouts where the model put a fighter at least 2 points above the opening price, and how far the price then moved towards the model by the close. Lines are 95% intervals.
On its own the model forecasts worse than the opening price. Blended in, it moves that price, posted a median of three days before the fight, 87% of the way to where it closes.
A forward test was frozen and published on 24 September 2026, before any bout it covers had been fought. It closes on 25 September 2029. Every result so far is a backtest.
Blended into the opening price, and tested with e-values
Four windows, each scored as an e-value against the opening price, with a bar of 20 (80 for 2023–2025, registered as one of four hypotheses), on a log scale. Two were pre-registered in public: for 2016–2020 the protocol’s hash was timestamped in Bitcoin blocks and pushed to GitHub before the model was trained; for 2026 the rule was pushed to GitHub before the holdout model was trained.
Paid client work, private.
A commissioned system that archives the full Zacks stock screen every trading day in a SHA-256 hash chain and turns it into daily recommendations for an equal-weight book of 25. The rules were frozen before the first one; since 1 October 2026 a paper book has filled every recommendation, none skipped.
Open source, MIT or Apache-2.0. A plugin for Claude Code, past 1,000 downloads.
A skill for Claude that tutors you through anything you are teaching yourself. Theory comes on a sheet you read and then close, drills are done on paper with a check beside every answer, and every new skill is tested again, cold, two days later.
The same day
Practice. It does not count yet.
Two days later, cold
Measurement. This score counts.
Open source, MIT.
Native clipboard history for macOS and Windows, written natively twice rather than once in Electron. No account and no network calls, and anything a password manager marks as concealed or transient is never written to disk.
brew install --cask romanprigodskii/tap/clipwellPersonal
Business
Built on an internship at Alfa-Bank. A concept, 93 screens.
The bank of 2035: one passport and as many profiles as a life needs, roubles and crypto in one balance, and an AI copilot that drafts an action and waits for a yes.
Every name, balance and date on these screens is demo data.
Two sole-authored papers, both grown out of evaluating Vertex MMA.
Accepted at two workshops
A poster at NewInML and at E-Values: From Statistics to ML, both NeurIPS 2026 workshops. E-Values accepted the shorter version.
I wrote down 84 hypotheses about where the market’s prices for UFC fights might be wrong and Vertex MMA right, before any of them was scored, and tested each with e-values against the prices actually posted, over 1,787 fights. The headline result is negative: none clears the bar for all 84 at once.
At fair prices, six of the 84 clear 20, the bar for one hypothesis.
Point at a mark, or use the arrow keys.
Working paper
How small an improvement can a model evaluation actually see? Retraining the same model under a different random seed changes nothing real, yet it reproduces 80% of the only improvement Vertex MMA’s winner model ever shipped.
Everything this pipeline measured on its main pool sits inside its own detection floor, including the one change it shipped.
Published, Young Scientist, February 2025
My MYP personal project, written in English in Grade 9 with my supervisor Kirill Kusakin: the residues a smallest counterexample would have to fall in, and other authors’ approaches. Above, the path of 27.
School project, Grade 9, 2024
Why the cut edge of the newspaper unrolls into a sine, proved and then measured: traced from the scan, the edge follows a sine to within 1.8% of its height.
School project, Grade 9, 2025
A Turkish tea glass traced with four parabolas and cut into ever thinner discs: 216.9 cm³, inside the 200 to 225 ml the water test gave.
The research page: both papers and the school projects, figure by figure
Olympiads, by school year.
Checked against the published competition protocols where they exist. The All-Russian finals had 91 finalists in Grade 9 in 2025 and 118 in Grade 10 in 2026; more than 470,000 students entered its economics olympiad in 2024/25.
Open to research and internships.
Based in Moscow. Applying for Fall 2027 entry.