Roman Prigodskii

The system

Not one model. A winner ensemble over 118 features (LightGBM, CatBoost and logistic regression, blended on a validation window and order-averaged over both fighter orderings, then a single post-blend age coefficient), a debut specialist, a method model, a debut method model, a cause-specific Poisson finish hazard, a no-intercept logistic decision model, and a ten-thousand-run Monte Carlo simulator. Live in English and Russian: 35 routes in two locales, 1,533 translation keys each, on Next.js 16, React 19, TypeScript 5.9 strict, Tailwind 4, Postgres and Python 3.12. Virtual currency only.

Measured against the market

Held out from January 2025, read directly from committed artifacts. The comparison is reported on every basis, including the ones where the closing line wins, because that is the comparison worth having: the model is better calibrated than the line and less sharp, so what remains is resolution, which needs information the public record does not contain.

Discipline

• Point in time, five ways. Snapshot-then-apply history, strictly-earlier rating lookups, a separate chronological ratings replay, a date-bounded pre-UFC career walk, and a frozen percentile-clipping anchor.
• The market is never an input. Odds are stored, shown and used for scoring, and never enter the feature matrix. Removing them cost 1.5 points of test accuracy. They stayed removed.
• A measured detection floor. Five seeds on a 3,087-bout walk-forward pool put the minimum detectable effect at 0.0036 nats; the one lever ever shipped to the winner leg is worth 0.0026 and shipped only because three independent legs agreed on its sign. The repository says exactly that.
• A written record of failure. Two dozen ideas built, gated and rejected, documented beside the ones that shipped. One shipped feature was later identified as a leak and removed.

What it is

One monorepo across Android in Jetpack Compose, iOS in SwiftUI, a web client, a Python and FastAPI backend, the infrastructure and the protocol documents. The cryptography is ephemeral key exchange, per-message key derivation, Argon2id and elliptic-curve signatures, with WebRTC peer-to-peer calls and a contextual AI assistant built in.

I co-founded it, I write its code, and as Head of AI I own everything the product does with a model, starting with the assistant that sits inside every conversation.

How it shipped

Nothing went to the stores without clearing a written release gate: compile green, 309 unit tests passing, lintVital clean on the release build, and R8 keep-rules verified against the 15 wire fields that have to survive minification. The Android 1.0.0 build shipped as .aab and .apk with mapping files and checksums retained.

Audited twice, adversarially

Two formal codebase audits are on record. The first, in April 2026, raised about 58 findings. The second, before release in August, raised 74 candidates and then ran an adversarial second pass over them: 42 were confirmed, and 32 were refuted and written down as refuted, so the next audit does not spend time on them again.

The thesis, and what happened to it

The project began from a claim: the UFC closing line is sharp, but regional and club boxing lines are soft, so that is where a fundamentals model would find room. Cut by level, the priced test set says the reverse. Club fights over four to six rounds return −20.2% at a fixed 2% edge threshold. Twelve-round title fights carry the strongest closing-line value, +0.0183, and +14.6%. Three independent markers of level, distance, belts and depth of record, all point the same way.

The mechanism is the useful part. The model reads records and ratings, and a rating is only as good as the record under it. At world level both fighters have thirty bouts behind them; on a club card the market knows which of the two is being managed, and the model is reading two thin records.

Checked on a window the rule never saw

The model was refit up to June 2021 and the rule copied, unedited, onto June 2021 to June 2023. The direction held: the top of the market gave 1.8 times the closing-line value of the bottom, +0.0270 against +0.0150, with barely overlapping intervals. The money did not: +4.2% with zero inside the interval, against +11.3% on the window where the cut was chosen. That is the shrinkage a post-hoc cut predicts, and the record treats it as part of the answer.

The leak it caught

Judges' scorecards are published only for fights that went the distance, so the mere presence of judge fields told the model how a fight had ended, a fact nobody has before the bell. Where judges were recorded, 9.7% of fights ended early; where they were not, 74.3% did. A control on 5,673 events, same page and same crawl, confirmed it was the outcome and not the crawler. The leak was worth +0.0042 [+0.0035, +0.0049] on the confirmation half. It is gone, and the bench was then made to reproduce itself exactly, after two identical runs were found to disagree by 0.0002.

Where it stands

As of 4 August 2026, and all of it backtest. On its own the model does not beat the closing line: the gap is −0.0211, or −0.0198 with a yearly refit. In the same run, blending it into the market price improves that price by +0.0043 [+0.0024, +0.0062]: it carries information the price does not. On twelve-round and continental or world title fights, 803 backtest bets show closing-line value of +0.0143 [+0.0104, +0.0182] and +8.7% [−0.0%, +18.0%], and the one interval that ever excluded zero no longer does. The margin paid at the open is about twice the closing-line value, so at the closing price the strategy's expectation is negative. The project says exactly that, first.

Scored as two-outcome log-loss: adding draws to the target was measured, and moved nothing.

What it does

The only project here that somebody else specified and paid for. It ingests screen exports from Zacks, turns them into target positions, reconciles those against what the Interactive Brokers account actually holds, and places the orders that close the gap. Collection runs on a schedule on a VPS.

Why dry run is the default

The failure mode is money. So the pipeline does nothing to the account unless it is told to, every run reconciles before and after, and a malformed export or a surprising position stops the run rather than guessing its way through it.

Written twice

The same bar, the same cards, the same keys and the same history format, implemented natively on both platforms rather than once in Electron. No account and no subscription. Installing it is brew install --cask romanprigodskii/tap/clipwell, which meant shipping a cask, a release pipeline and a product site alongside the app itself.

The detail nobody sees

Content that a password manager marks as concealed or transient is never written to disk. Clipboard managers are a quiet way to leak secrets, and this one declines to.

The bank of 2035

In 2035 your bank is also your mobile operator, your accountant and your crypto exchange. One passport opens as many profiles as a life needs: personal, business, a budget shared with a spouse, a child's card. Roubles, the digital rouble and crypto sit in one balance, because by then nobody should care which rail the money took. Even the cashback is flexible: it can arrive in gigabytes.

What actually runs

The rates are live, from the Bank of Russia's daily fixing and from CoinGecko. The copilot is Claude behind guardrails: it drafts the action and waits for a yes, and when there is no key or no network a deterministic stand-in takes over, so the demo never falls over halfway through a pitch. The wallet is a closed economy on Postgres: sign up with a phone number and send play roubles, or play crypto, to another phone.

A loan that explains itself

The credit screen does not stop at yes or no. It shows the three factors behind the limit, then lets you ask what if: close that credit card, and watch the limit recount.

The business half

Romeo Business lives in the same app: accounts with crypto treasuries, acquiring by payment link, QR, NFC and terminal, stablecoins converted to roubles on arrival, a team where a large payment waits for a second signature, and an AI accountant that draws the next ninety days of cash flow and points at the gap two weeks before it opens.