M.A. Quantitative Methods in Social Sciences, Columbia University
(Major GPA: 4.0 · A+ in NLP & Machine Learning · Overall: 3.92)
Targeting roles in London · Zurich · Geneva.
I build time-series models, Bayesian inference pipelines, and NLP systems applied to financial markets, and the local, governed AI tooling around them, on a 16 GB laptop. Previously Trade Analyst at Marex (energy derivatives, PnL tracking, sentiment tooling for morning reports). Certified NFA Series 3.
| 📡 AI-Infra Model Drift | Walk-forward drift monitor on 10 AI / semiconductor equities — SARIMAX · XGBoost · LSTM with PSI / KS / Page-Hinkley / rolling-RMSE detectors across 4 macro regimes (2014–2026). Built to measure the sentiment-return drift the thesis below first observed. |
| 🔬 FinBERT × SARIMAX | M.A. thesis — earnings-call sentiment vs renewable energy returns. Sentiment coefficient sign inverted pre/post-COVID (+79 → −90). |
| 🛡️ AgentOS Control | macOS menu-bar app that governs the coding agents of my AgentOS runtime: a risky tool call becomes a notification, and Approve requires Touch ID. The notification appears 0.5 s after the request; every decision leaves a hash-chained receipt. SwiftUI, 70 tests. |
| 🧠 MiniMind Lab | A 64M-parameter LLM trained from scratch: CPU and Apple-GPU smoke tests on an M1 Pro, then 8 h on Kaggle's free 2×T4 (SFT loss 2.22 → 1.41), served by Ollama at ~540 tokens/s. In tokens per second, the two T4s run 37× the laptop CPU. |
| 🎓 OpenMAIC Quant | My own projects (drift detection, the thesis, Bayesian coursework) turned into revision courses by an open-source multi-agent classroom, run locally: 8 scenes from 21 LLM calls, every number on the slides checked against the source project. |
| 🍷 Wine Recommender | Constraint-satisfaction chatbot. Relaxation fallback raises coverage 27% → 66% — same logic as portfolio optimization under active constraints. |
| 🎬 Dynamic Video Generator | Fully local multimodal gateway — one same-origin proxy in front of LTX-Video, Draw Things and Ollama. Budgets work in pixel-frames rather than seconds, because that is what decides whether a 16 GB machine swaps: 12 s/step at 35 M against 104 s at 221 M, a ≈9× cliff rather than a gradual slowdown. Interactive architecture. |
| 📱 MiniVid | The same idea on a phone but offline: a photo and a prompt become a short clip, generated locally on an iOS device. Core ML Stable Diffusion frames, interpolation, camera moves, a recorded voiceover and an mp4 export. |
Python · R / Stan / brms · statsmodels · SARIMAX · FinBERT · BERT · scikit-learn · XGBoost · PyTorch · SQL · PowerBI · Tableau · FastAPI · Flask · pandas · SwiftUI · Claude Code · MCP · Ollama
📫 [email protected] · LinkedIn · France — open to relocation