Data Science · AI · Building in public
Data Science & AI graduate from Le Wagon, building practical projects in public — containerised ML APIs, AI agents, automations. My 10 years in digital marketing help me frame the right problems and ship measurable results.
01
After 10+ years running digital marketing, I specialised in Data Science and AI through Le Wagon’s intensive bootcamp. Today I build hands-on projects — FastAPI services deployed on GCP Cloud Run, prediction models, multi-agent systems — open-sourced on GitHub.
How I build: start from a clear business problem, compare a baseline to an improved approach, document assumptions and limits, ship reproducibly. Reproducibility, measurability, honest reporting of trade-offs — for projects that matter, not demos.
My marketing background is still an asset — I read data with a business lens and I know what a KPI means to a team. Goal: a first role as a junior Data Scientist, ML/AI internship, or collaboration on a real-world data/AI project — French-speaking Switzerland, hybrid or remote.
02
03
A selection of personal projects, live or in active development.
ML prediction service deployed on Google Cloud Run: containerised FastAPI app with a prediction endpoint, interactive Swagger docs and a reproducible project structure (Dockerfile, Makefile, tests). Designed as a reusable backbone for exposing an ML model in production.
Multi-agent AI system that automates the job search: evaluates job offers on multiple dimensions, generates ATS-optimised CVs and submits applications via Playwright.
NVFP4 (compressed-tensors, ~18 GB) checkpoint of Qwen3.8-27B, packaged for vLLM on NVIDIA Blackwell / DGX Spark (GB10). Weight-space edit (abliteration on layers 18–51, not a fine-tune), then oneshot quantization with llm-compressor. Aimed at local Spark inference with up to 262k context and an FP8 KV cache.
Bilingual (FR/EN) evaluation set to measure LLM refusal vs compliance after abliteration. French translation of user prompts (GX10 qwen-coder, translate-only), Qwen3Guard labels, no assistant completions — not a fine-tune corpus. Deduplicated and published on Hugging Face.
Minimal MCP server for reading and sending emails via IMAP/SMTP — zero external dependencies (Python stdlib only). Compatible with Claude Desktop, Cursor, Hermes and any MCP client. Works with any mail provider (Infomaniak, Gmail, iCloud, Proton Bridge…).
Benchmark tool measuring LLMs' ability to accurately recall specific code in long contexts: tests positional recall by asking models to reproduce JS/Python function definitions verbatim, with automated line-by-line scoring and comparative Plotly dashboards.
04
A strong marketing background, rounded off by an intensive Data Science & AI bootcamp.
Python, ML, deep learning, applied AI in enterprise.
05
A decade at the crossroads of digital marketing, data and operational leadership.
05
Open to a first role as junior Data Scientist, ML/AI internship, or collaboration on a real-world data/AI project. Drop me a line.
La Tour-de-Peilz · French-speaking Switzerland