Juan Mediavilla · MSc Machine Learning, UCL · London

I build research systems end to end.

The data platform, the models, and the evaluation that decides whether they are worth anything. That last part is the one I would want to be judged on.

  • ~1.05 TBof market data collected and archived, as of Aug 2026
  • 188,856trades in the licence-clean audited panel
  • 34/34markets improved by pooling across markets
  • 718commits on the research platform, built solo

Selected work

Four builds.
Each one has its own page,
with the results.

Built end to end, then measured

01

Prediction-market research platform

Six scheduled collectors and real-time ingest feeding an ~1.05 TB archive, with live inference, order execution, and a backtester that simulates realistic fills. 718 commits
400-market panel
02

Latent market dynamics — MSc thesis

A model of when prediction markets move rather than where they go, fitted with a Metropolis-within-Gibbs sampler I wrote from scratch. custom MCMC
34/34 markets
03

Market-data portal

A seven-stream archive turned into a documented, versioned data product, with provenance enforced by the build rather than by a promise. 7 streams
12.8 GiB/day
04

Activity recognition — BSc thesis

A full self-supervised pipeline benchmarking six architectures on wearable-sensor data, with every model tested on people it had never seen. 6 architectures
23-min baseline held

All projects →

Approach

Build the system, then find out whether it works

A 24/7 market-data platform feeding a live inference and execution pipeline. A Bayesian sampler written from scratch because the standard tools did not apply. An on-chain wallet-intelligence engine. A fund-accounting system. A six-architecture deep-learning benchmark.

In each case I built the thing, and then built the test that could have shown it was worthless. That second half is the part I would want to be judged on. It is also why I can tell you which of these results I trust, and how much.

A result that goes against you is a credential. It is the evidence that the rest of the numbers were not chosen after the fact.

How I work, with the artifacts →

Contact

Available from September 2026, in London

Looking for machine-learning engineer, applied scientist, or quant research and development work at a place where the results have to survive someone checking them.