Data Scientist · CS Student · CESI Orléans × LAB'IA Loire Valley
1.0 ┤╮
0.7 ┤╰─╮
0.4 ┤ ╰─╮ ╭ val
0.1 ┤ ╰─╯──── train
└──────────────────
fit() ✓ overfit: low
CS student @ CESI Orléans, working as a Data Scientist (work-study) @ LAB'IA Loire Valley. I'm interested in understanding how models work and building systems on top of them.
Outside of code: geopolitics, philosophy, and future of AI. Sometimes I contribute to open-source software, and I'm big on editing Wikipedia (17k+ edits).
x ──▶ [embed]
│
[attn] ◀── Q K V
│
[ffn]
│
[out]
Designed and deployed an internal multi-agent LLM chatbot (20+ constant users) to streamline workflows and knowledge access via RAG. Set up local server infrastructure and security for future production deployments.
Built GenealogyAI — an autonomous genealogy agent for French departmental archives. Enter an ancestor's name: the agent navigates Arkotheque (~80 departments), OCRs handwritten 19th-century birth records with Gemini 3, extracts parents, and climbs the family tree generation by generation. If an ancestor is born after 1900, the app auto-generates the mairie request letter.
AgriTech challenge: early detection of sick livestock to speed up care. Proposed a rail-mounted thermal camera system scanning enclosures, detecting temperature anomalies per animal, and alerting the farmer in real time.
Sustainability challenge: extending the life cycle of NFC-enabled paper transit tickets. Most users discard them unaware of the reusable copper trace inside. Proposed anti-vandalism collection bins at stations and a depot-side processing machine validating ticket reusability via NFC read + computer vision (physical condition check).
On-premise LLM chatbot built at LAB'IA Loire Valley for businesses seeking local AI solutions. Multi-agent orchestration (LangGraph), local models (Ollama), and RAG over company documents — no cloud dependency.
"user query"
↓ tokenize
[1024][421][89][1337]
↓ embed
[0.23, -0.41, 0.87 …]
↓ attend + retrieve
↓ generate
Mobility gap analysis for Orléans' public transport network. Building a GeoJSON time-series dataset (572 bikes, every 5 min) to cluster usage patterns and detect coverage gaps vs existing bus/tram schedules. Stack: Python · GeoPandas · time series · clustering.
Architectural similarity score between LLMs. Quantifying how "close" two models are based on their internal structure — attention patterns, layer config, parameter distribution.
Web app for tracking internship applications. PHP/Symfony architecture, deployed on Azure VM.
Fine-tuned Qwen3-0.6B with LoRA for automatic conversation title generation in French. Synthetic dataset created via Ollama from OpenAssistant/oasst2 (~2000 examples).
Statistical analysis of anonymous IP contributions on French Wikipedia. 646 edits collected via automated hourly pipeline, classified by vandalism type, and analyzed across temporal, geographic, and ISP dimensions. Presented at the Wikipedia Patrol community meeting (Jan. 2025).
Firefox search engine extension — adds a custom @genius keyword to the address bar
for instant Genius lyrics lookups without leaving the browser.
↑
│ king• •queen
│
│ man • •woman
└──────────────────→
gender dimension