Alexandre Dalban
Portrait of Alexandre Dalban

Alexandre Dalban

Now
Based in London. I finish my MSc in Artificial Intelligence at CentraleSupélec, Paris, in November 2026
Before
Until 16 October I was an applied AI research intern at Siemens in Munich. Before that I did a BS in Mathematics and Computer Science at Boston College, and spent two summers doing alpha research at LeCap Asset Management, a systematic equity hedge fund in London
Next
I'm looking for full-time roles from November 2026 in London, the US, the EU or Switzerland. I'm a US and EU citizen with UK settled status, so I don't need visa sponsorship in the UK, the US or the EU
Contact
alexandre.dalban@gmail.com · GitHub · LinkedIn · CV
Start with
the Polymarket trading system, seven months of strategies with two postmortems and the wallet curve, and Phantom, a forecaster I audited in July and corrected in the README

Work

  1. 2026
  2. Sepproject Notes for quant and AI interviews published A free 315-page reference I wrote while interviewing, covering probability through market microstructure, with 191 worked problems. Fifty-seven of them are questions I was actually asked in 2026 assessments and interviews.
  3. Mar →project Fantasy Premier League, predict and optimise live season A LightGBM point predictor with a minutes model feeds a mixed-integer optimiser that plans transfers three gameweeks ahead. Chip-free replays of the last three seasons scored 2,129 to 2,467 net points. I'm now using it to plan a real entry each week and tracking that on the page.
  4. Julhackathon Counterproof, a venture-diligence engine demo shipped It scores founder claims by the evidence that should exist if they were true, and prices absence with a likelihood ratio over an append-only ledger. Team of three, 21-hour sprint. There's a static demo on fixture data.
  5. May–Octwork Applied AI research intern, Siemens AG, Munich done I worked on a step CAM engineers still do by hand, deciding which faces of a CAD part form one machining feature. On real industrial parts, a face-pair classifier went from 51% to 76% of features recovered exactly, mostly through better labels. On a separate set of 119 held-out parts, an LLM agent inside Siemens NX with nothing trained for the task got 77%, against 61% for the classifier. The write-up is my MSc thesis.
  6. Aprresearch Phantom, distributional forecasts for 362 cryptocurrencies audited It's a 31.7M-parameter Transformer that predicts Student-t distributions, with an out-of-sample rank IC of 0.124 at 10 days, positive in 15 of 15 months. I audited it in July and found the Sharpe I'd reported was inflated by about √10, so I corrected it in the README and left the old numbers up as errata.
  7. Aprcompetition IMC Prosperity 4 done Our team of three finished at rank 458 of 18,803 teams (Finalist Crew). A market simulator, written in Rust with Claude Code, reproduced the round-1 bots' quoting rules at 99 to 100% exact match, and there are five documented dead ends.
  8. Dec 25–Jun 26project A live trading system on Polymarket wound down My collaborator and I built it and ran it on personal capital for seven months, through two generations and a rewrite. The first priced daily crypto markets off the Deribit options surface. The second pressed the favourite on 15-minute markets, which made 3,037 trades in 18 days at a 78.0% win rate. It traded $260k of volume across 6,213 markets to finish $155 up, and there are two postmortems.
  9. Febhackathon DeFi Sentinel, HackEurope 2nd place We took 2nd place in the Solana challenge with a rug-pull detector for new Solana tokens. It was a team project, and I built the live-data collectors and the first backend. My teammate's best model (AUC 0.9995) drew 89% of its importance from deployer history, which doesn't exist when a new token is scanned, so we shipped the restricted one instead.
  10. Febdataset taut-arb-backtest, an open dataset public The repo holds 30,180 settled Polymarket crypto markets with 1.1M price observations, plus 5,171 sports events aligned to de-vigged Pinnacle odds. It's data only.
  11. Oct 25–Apr 26research A domain-specific LLM for reliability engineering done We fine-tuned Qwen3-8B on 866 verified question-answer pairs bootstrapped from 56 textbook problems. Under 10-fold cross-validation it went from 63.6% to 82.2% (p = 0.002), but that's in-distribution, and on a 54-question holdout it did worse than the base model. It's a joint project.
  12. 2025
  13. 2024–25work Quantitative finance intern, LeCap Asset Management, London two summers I validated 100+ alpha signals for a systematic equity strategy and built the weekly reporting that scores 150+ signals on 8 metrics. The best forecasting model reached 54% daily directional accuracy, against a coin-flip baseline near 50%.
  14. 2024
  15. Sep–Deccoursework TensorTalk, zero-shot voice cloning done Training-free voice conversion from WavLM features, with nearest-neighbour matching and a HiFi-GAN vocoder. Team of four at Boston College. I'm first author on the write-up.