Production systems delivered at Téchnéos, plus personal and academic work. Professional case studies are written without client names, product names or customer data — each page describes the engineering and labels which numbers are measured and which are estimates.
A reusable, self-hostable foundation for enterprise AI applications: identity, fine-grained authorization, AI gateway, governed tool execution, durable workflows, audit and observability. Nine milestones in ten weeks, with architectural invariants enforced in CI.
Digital-twin and design-of-experiments platform for industrial process and formulation R&D. Includes a distillation soft sensor reaching R² 0.81 with 90% prediction intervals that cover 94–97% of held-out measurements.
Probabilistic demand forecasting with quantile regression, driving automated safety stock and restocking decisions under a target service level across 300+ SKUs. Four model families compared per series.
Indexes a manufacturer's PFMEA and 8D quality documents, then searches, chats, cross-references and drafts from them — ~2,090 records, trilingual UI, running entirely inside the customer network on local models.
Remote GPU-accelerated OCR plus schema-validated LLM extraction, turning dense technical documents into standardized datasets — with a visual schema editor and a multi-run evaluation benchmark.
Supplier invoices in, ISO 20022 payment file out: OCR, LLM extraction, master-data validation and a human supervision desk for everything uncertain. Crash-resumable workflows, full audit trail, no automatic money movement.
Work from my MSc in Data Science and personal research. Code for these is public on GitHub.
Quantitative finance
Transition matrices built from EMA and RSI regimes over twenty years of equity data, with expected-value scoring and threshold-driven trading — benchmarked honestly against buy-and-hold, which outperformed it.
Recommender systems
Collaborative filtering and content-based recommendation over 3.6 million user interactions from the Million Song Dataset — matrix factorization, implicit feedback modelling and hybrid strategies.
Predictive modelling
Predicting churn from usage, contract and demographic data with Random Forests and neural networks, identifying at-risk customers so retention effort goes where it changes the outcome.
Computer vision
CNN baselines, EfficientNetB0, MobileNetV2 and Xception compared on leaf disease images; Xception reached 98.75% accuracy, with batch size the decisive constraint against GPU memory.
LLM application
An AI dungeon master built on locally served Mistral and Llama 2 models. The interesting part is the scaffolding: how prompts and state carry combat, character creation and continuity between turns.
NLP · Forage programme
Sentiment analysis of airline customer reviews, from web scraping through NLP to predictive modelling of flight satisfaction.
Analytics · Forage programme
Exploratory data analysis and K-Means clustering over customer demographics, transactions and behaviour to find the segments with the highest revenue potential for a bicycle retailer.
Data engineering
A workflow automation system for collecting, updating, processing and reporting business intelligence data, architected around large datasets, periodic updates and readable report generation.