Master's thesis 2026. Synthetic health data.

The thesis measures the trade-off between differential privacy, utility and fairness for five families of synthetic data generators (AIM, MST, DP-CTGAN, PATE-CTGAN, DP-TVAE), on MIMIC-IV and UCI Diabetes 130-US.

  • Four axes: utility, statistical fidelity, membership inference attacks, subgroup fairness.

  • Bootstrap intervals and Holm-corrected Wilcoxon tests.

  • DP-SGD written from scratch, with zCDP accounting.

See the thesis code
Gaussian mechanismx̃ = x + 𝒩(0, σ²)
ε 1.98 · σ 2.44

Land or Water. The world map according to 24 LLMs.

24 open-weight models answer “land or water?” for 16,200 coordinates, with no image and no tools. I redraw their map and score it against a 1 km land mask.

See the benchmark code

Two models in the lead

MiniMax M3 at 88.8%, DeepSeek V4.1 Flash at 88.5%.

8 models biased towards “Land”

They do not beat “always Water”, at 71.0%.

One provider per model

Pinned through OpenRouter, with no silent fallback.

389,196 requests, $16.85

Read the dataset

Creative agent for Reels and stories

A one-line brief, such as “knitting app, cozy vibe, beginner audience”, comes in through the command line or the Flask API. A single call to Claude writes N structured JSON prompts (subject, props, photography, setting), passed as they are to Gemini, DALL·E or Replicate. Sora 2 can turn each image into a video.

Private code, demo on request

Profile

  • Master

    MSc Data Science for Business, École polytechnique × HEC Paris, 2024 – 2026

  • Bachelor

    Computer Science, Università Bocconi, Milan

  • Fields

    LLM agents, probabilistic ML, data science

  • Open to

    Freelance projects, applied AI and ML engineering roles

Portrait of Solal Zana

Services. End to end, or on one specific piece.

  • Agents and automation

    LLM agents and workflows plugged into your tools, which take repetitive work off your hands.

  • Models and data

    Forecasting, scoring and matching, rigorously evaluated, and synthetic data when the real data is sensitive.

  • Product MVPs

    From prototype to a first live version: API, Postgres or Supabase database, web or mobile app.