Roman Prigodskii

Letovo School, Moscow

Grade 11, final year. Russian national curriculum, advanced (profile) mathematics track. GPA 6.38 / 7.0, above group average in algebra and in every core quantitative subject inside the school's most competitive quantitative group.

SAT Mathematics 790, August 2026. A further sitting scheduled for December 2026. IELTS 7.0.

All-Russian Olympiad in Economics

National finalist in two separate years. The All-Russian Olympiad is the national academic competition system; reaching its final twice places him in the national cohort in the subject.

All-Russian Olympiad in Mathematics, regional stage

Grade 9. Protocol-verified.

Leonhard Euler Olympiad, regional prize-winner

Grade 8. The Euler Olympiad is the All-Russian mathematics olympiad at Year 8: same regional dates, problems set by the same central methodological commission. Read with the row above, that is a regional prize in two consecutive years.

All-Siberian Open Olympiad

Grade 7. Protocol-verified.

Letovo School, best MYP personal project in Mathematics

Director-signed.

Testing by betting when the bets are real

An e-value audit of 84 pre-registered market-efficiency hypotheses. Sole author. Submitted to the NeurIPS 2026 workshop E-Values: From Statistics to ML and to NewInML.

• The null is a bookmaker's posted price, the stake is a stake, and the wealth process is the e-process. 4,141 walk-forward out-of-fold bouts, 1,787 of them carrying an archived closing price.
• The 84 slices overlap, so Benjamini-Hochberg's guarantee is not established. Its four discoveries on the 1,191-bout discovery window, one favourable and three against the model, survive neither Benjamini-Yekutieli nor e-BH.
• A segment written off at p = 0.30 is still growing on the wealth scale: not refuted, 100 bouts short at fair odds.
• A post-hoc threshold rule is charged for the search it ran, the direction it could have searched, and the bookmaker's margin. The ladder is the finding.
• Validated before use: 40,000 null replications give E[e] = 0.985 ± 0.018 per segment. The headline result is negative and is reported as negative.

PDF, E-Values version  ·  PDF, NewInML version

Measure the instrument first

Detection floors for model selection and for LLM judges. Sole author. Submitted to the NeurIPS 2026 workshop TAE: Can We Trust AI Evaluation?

• The null lever: the same recipe refit under a different random seed changes nothing real, so everything it produces is instrument noise.
• On a 3,087-row walk-forward pool under five seeds, re-seeding alone reproduces 80% of the only lever ever shipped; the pipeline had 56% power against its own shipped result.
• A floor belongs to the instrument that measured the effect. Recomputed on the right pool, the project's one large win is 4x its own floor rather than 13x another pool's, correcting the paper's own earlier draft in print.
• The protocol transplants: on a deployed LLM classifier, two identical reruns disagree on 5.6% of items and the floor is 2.9 points of agreement.

PDF

Vertex MMA, vertexmma.com

UFC analytics and forecasting platform, English and Russian, virtual currency only. Sole developer.

• 4,581 fighters, 8,906 bouts, 795 events. 243k score-history rows, 86k career bouts, 49k ranking snapshots, 41k round-stat rows, 31k judge scorecards.
• Nine models including a 10,000-run Monte Carlo simulator: a 118-feature winner ensemble over LightGBM, CatBoost and logistic regression, a debut specialist, method models, a cause-specific Poisson finish hazard, a no-intercept decision model.
• Held out from January 2025: accuracy 0.6747, log-loss 0.6137, Brier 0.2124, AUC 0.7244 over 664 bouts.
• On the 582 priced bouts, 0.6753 / 0.6171 / 0.2140 against the closing line's 0.6838 / 0.5922 / 0.2035: better calibrated, less sharp, and the remaining gap is resolution, which needs information the public record does not contain.
• Point-in-time enforced by five independent mechanisms. Bookmaker odds are never a feature: removing them cost 1.5 points of test accuracy and they stayed removed.
• Roughly two dozen rejected ideas documented beside the shipped ones, each with its reason.
• Next.js 16, React 19, TypeScript 5.9 strict, Tailwind 4, Postgres, Python 3.12. 35 routes across two locales, 1,533 translation keys each.

Source

Gluline, gluline.com

End-to-end encrypted messenger, released on the App Store, Google Play and the web. Android in Jetpack Compose, iOS in SwiftUI, a web client, Python and FastAPI backend, infrastructure, protocol. Co-founder, engineer and Head of AI.

• Ephemeral key exchange, per-message key derivation, Argon2id, elliptic-curve signatures, WebRTC peer-to-peer calls, built-in contextual AI assistant.
• The Android 1.0.0 production build shipped as .aab and .apk, with mapping files and checksums retained.
• Release gate at the August 2026 audit: compile green, 309 unit tests passing, lintVital clean, R8 keep-rules verified.
• Two formal codebase audits on record. August 2026: 74 candidate findings narrowed to 42 confirmed by an adversarial second pass, 32 written down as refuted.

Zacks to Interactive Brokers execution pipeline

Commissioned and paid third-party work. Screen ingestion from Zacks exports into an Interactive Brokers execution path, VPS hosted, scheduled collection, dry-run by default.

Clipwell

A native clipboard history for macOS (Swift, SwiftUI) and Windows (C#, WPF). MIT licensed, distributed as a Homebrew cask, English and Russian UI. Content marked concealed by a password manager is never stored.

Product site  ·  Source

Vertex Boxing

The Vertex method on professional boxing, built to test whether regional and undercard markets are priced softly enough to leave room. They are not, and the project records that as its finding.

• Cut by level, club fights over four to six rounds return −20.2% at a fixed 2% edge threshold, while twelve-round title fights carry the strongest closing-line value, +0.0183. On a window the rule never saw, the top gave 1.8 times the closing-line value of the bottom.
• Blended into the market price, the model improves it by +0.0043 [+0.0024, +0.0062]; on its own it does not beat the closing line, and the margin paid at the open is about twice the value it finds.
• Caught and removed a post-bell leak: judges' fields exist only for fights that went the distance, so their presence told the model how the fight ended (early finish 0.097 with judges, 0.743 without).
• All backtest, as of 4 August 2026. Zero real bets placed.

• Measure against the hardest available baseline, and publish the comparison on every basis, including the windows where the baseline wins.
• Gate before shipping: every change is measured against a pre-computed detection floor, and the one Vertex lever that shipped below its floor did so because three independent evaluation legs agreed on its sign.
• Keep the rejects, with the reason each was rejected.
• Report negative results as negative, in the abstract and not in a limitations section.
• Claim precisely. Submitted means submitted. An award cites its protocol. A number carries the window it was measured on.

Python, LightGBM, CatBoost, scikit-learn, NumPy, pandas, statsmodels, Monte Carlo simulation

E-values and test martingales, anytime-valid inference, walk-forward validation, calibration and sharpness decomposition, power and detection floors

TypeScript, Next.js, React, Tailwind, Postgres, Supabase; Kotlin and Jetpack Compose, Swift and SwiftUI, C# and WPF; Python and FastAPI, NestJS, Docker, Linux, Caddy, SOPS

Russian (native), English (working). Bilingual technical writing. LaTeX.