About
Juan Mediavilla
I study MSc Machine Learning at University College London, and I build research systems end to end — the data platform, the models, and the evaluation that decides whether they are worth anything.
I want to be a machine-learning engineer or applied scientist, in research or quantitative finance, at a place where results have to survive someone checking them. Spanish, based in London, available from September 2026.
Experience
What I have built, and what it shows
No employer has paid me to build these. Every one is either academic work or something I built myself, and each has its own page with the results.
-
Prediction-market research platform
Data engineering and applied ML at production cadence: six collectors on
systemd timers, real-time ingest, live inference and order execution, and a
walk-forward backtester whose pre-registered verdict was that none of the four
strategies beat random.
2026 – ongoing
718 commits -
MSc thesis — latent market dynamics
Bayesian modelling from first principles: a hierarchical point-process model
fitted with a Metropolis-within-Gibbs sampler I wrote from scratch, because no
standard tool applied.
2025 – 2026
UCL -
Wallet-intelligence engine
Anomaly detection under a high standard of proof: three independent tiers must
agree before an account is labelled, and the engine caps its own conclusion when it
cannot trace funding.
2026
51 tests -
Fund-accounting terminal
Financial systems correctness: NAV unitisation, decimal-exact money, and one
ledger shared by a command-line tool and a web terminal so the two cannot disagree.
2025 – ongoing
4 generations -
Market-data portal
Product engineering with provenance enforced by the build — including retiring
the product's own original selling point once I could not support it.
2026
7 streams -
BSc thesis — activity recognition
Deep learning and experimental design: six architectures benchmarked under
leave-one-subject-out validation, with a Friedman test on the result rather than a
leaderboard.
2025
Leiden -
Teaching Assistant, Leiden University
Two computer-science courses in one semester, one of which recruited me
personally. Labs, tutorials, grading and assignment design.
Autumn 2024
2 courses
Education
Two degrees
University College London
Thesis: Latent Dynamics of Prediction-Market Price Paths, supervised by Prof. Philip Treleaven. A hierarchical point-process model of when prediction markets move, fitted with a Metropolis-within-Gibbs sampler I wrote from scratch, and benchmarked against self-exciting and neural alternatives. Write-up →
Coursework: Supervised Learning · Probabilistic & Unsupervised Learning · Bayesian Deep Learning · Approximate Inference · Statistical Learning Theory · Reinforcement Learning · Information Retrieval.
Thesis: Fine-tuning General Time Series and Accelerometer-Specific Self-Supervised Models for Human Activity Recognition. Built a full self-supervised learning pipeline and benchmarked six architectures on wearable-sensor data under leave-one-subject-out validation. Summary →
Minor in Technology-Based Entrepreneurship at TU Delft, where I solo-built the web application for a five-person final project.
Teaching
Teaching Assistant — Leiden University
Human–Agent Interaction, and Concepts of Programming Languages (OCaml and Prolog). I was recruited to Human–Agent Interaction personally by its instructor, as a top-scoring student on that course the year before.
The programming-languages cohort was roughly 120 to 150 students. I graded 20 OCaml group submissions, 14 Prolog group submissions and 6 retake students; ran labs and tutorials; and co-designed assignments, including calibrating question and coding difficulty.
Recorded feedback described me as "a fantastic TA — proactive, engaged, and helping improve the course." Both course professors wrote my Cambridge and UCL recommendation letters.
Skills
What I can demonstrate today
Tiered honestly. What follows is what I can show working code or committed results for. What I have only studied is listed separately below, and labelled as such.
- AI tooling and agentic systems
- My core strength: designing agentic pipelines, writing the contracts agents operate under, and building the guardrails that make their output trustworthy — provenance requirements, constraints enforced in code rather than in prompts, and independent verification of what an agent claims it did. The published version is default-fail, which is MIT and readable in full. How I work, with the artifacts.
- Experimental design and statistical validation
- Specifications committed before outcome statistics; clustered bootstrap with the correct resampling unit; leave-one-out cross-validation over subjects and over markets; Friedman tests; Benjamini–Hochberg correction; walk-forward out-of-sample validation; recovery gates that must pass on simulated ground truth before a real result is read.
- Python
- Primary language.
- Applied ML in production
- XGBoost with isotonic calibration — a method that makes a model's stated confidence match how often it is actually right — plus live inference and order execution. Backtesting discipline: realistic fill simulation, walk-forward evaluation, realized-at-resolution profit and loss, and honest abstention when a required signal is absent from the data rather than substituting a proxy.
- Point processes
- Point processes are the mathematics of when things happen in time rather than how large they are. I work with two: self-exciting models, where each event makes the next more likely, and hidden-mode models, where the system switches between a quiet and a busy state you cannot observe directly. Hawkes and Markov-modulated Poisson processes at thesis depth — hierarchical partial pooling with a custom Metropolis-within-Gibbs sampler and a conjugate Normal-Inverse-Wishart population block, benchmarked against Hawkes and EWMA baselines.
- Self-supervised and representation learning
- Self-supervised learning trains a model on a large pile of unlabelled data first, then fine-tunes it on the small labelled set you actually have. Fine-tuned SimCLR, SelfPAB and the MOMENT time-series foundation model.
- Data and systems engineering
- Real-time WebSocket pipelines, Linux and VPS operations, systemd, Git, rclone. A 24/7 collection pipeline on a 15-to-60-minute cadence with gap auditing and continuity accounting.
- Web and front end
- React, JavaScript, HTML, CSS, Tailwind, Astro — four shipped front ends across three idioms: a single-page app, a Next.js application with third-party auth, and zero-dependency static (this site).
- Other languages
- Java, OCaml and Prolog — the latter two to teaching standard.
- Markets and microstructure
- Market microstructure, Kelly sizing, options basics, resolution mechanics, prediction-market structure, NAV (net asset value) unitization and decimal-exact ledgers.
- LaTeX
- Thesis writing.
Models I have implemented, trained or fitted
| Model | Where |
|---|---|
| XGBoost with isotonic calibration | Live inference and order execution on the research platform |
| Hierarchical Markov-modulated Poisson, custom Metropolis-within-Gibbs sampler | MSc thesis — written from scratch, no autodiff gradient available |
| Hawkes self-exciting process | MSc thesis — the model that beat my own |
| EWMA intensity baseline | MSc thesis — the baseline everything had to clear |
| Hierarchical Bayesian regime model, fitted by ADVI | Earlier market-regime work |
| SimCLR contrastive pretraining | BSc thesis — highest scorer of the six |
| DeepConvLSTM | BSc thesis — supervised conv-recurrent hybrid |
| SelfPAB · SelfHAR | BSc thesis — accelerometer-specific self-supervised models |
| MOMENT time-series foundation model | BSc thesis — fine-tuned for activity recognition |
| UMAP + HDBSCAN + Gaussian mixtures | Behavioural clustering over 3,413 wallets |
Studied, not claimed
Bayesian and variational inference, reinforcement learning, sequence models, statistical learning theory and information retrieval are coursework, not current capability. I have used several of them in anger, but they would need a serious refresher before I would claim them in an interview.
Measure-theoretic probability and stochastic calculus are an acknowledged gap, and I am not going to pretend otherwise. That is the main reason I am not currently positioning for pure theoretical-quant roles.
Looking for
What I want next
Machine-learning engineer, applied scientist, quant developer, or quant research at a shop where the work is empirical rather than theory-first. I am most useful where a system has to be built end to end and the results have to survive someone checking them.
Available from September 2026, based in London. My CV is here, and my email, GitHub and LinkedIn are at the top of this page.