§ Projects
Every project starts with a question. Three projects linked to the offers, and one research project. The dots show which stages of Question → Data → Model → Product each one covers.
A · Understand Completed Growth of an e-commerce marketplace +139.9% growth analysed A · Understand Completed Iphigen website, solar for SMEs 3 months from mock-up to launch B · Save time Completed Multilingual customer review analysis 3 languages FR · PT · EN Research Completed Will They Wake Up — EEG 1.5 TB of EEG explored
Python · SQL · Power BI
A · Understand · Completed
How can a marketplace support its growth while improving deliveries and customer experience? Context Brazilian marketplace Olist: 96,211 delivered orders, thousands of sellers, +139.9% growth over the period. Method 26 hypotheses tested: quality checks, corrected statistical tests, SQL, Power BI project. Result 20% of sellers make 82.2% of GMV. 6.8% of orders are late, costing 0.57 points of satisfaction. A measurable action plan. Express.js · FR/EN/ES
A · Understand · Completed
How do you show an Ivorian SME what solar would change on its bill, before any meeting? Context Iphigen installs turnkey solar systems in Abidjan using certified refurbished European panels. Method Figures from the client’s business plan, savings simulator, logo, mock-ups, copy, FR/EN/ES versions, audit form, Express.js development and launch. Result A website delivered in 3 months. Visitors drag their monthly bill and see their savings per month, per year and over ten years. Python · local NLP · ONNX
B · Save time · Completed
How do you understand what thousands of customer reviews say without reading them one by one? Context Beyond a few hundred reviews, reading by hand no longer works. The project extends the Olist study with a tool that can be reused on other data. Method Multilingual XLM-T model run locally (ONNX int8), “Undetermined” abstention below a 0.60 score, editable theme words and phrases. Result A tool that imports CSV, Excel, JSON, DOCX or PDF, flags cases to review and exports corrections. Online demo on 36 fictional reviews. Deep learning · LSTM · GCP
Research · Completed
Will They Wake Up — Deep learning on EEG signals Context After cardiac arrest, classifying patients’ neurological outcome from their EEG signals (“Good” label versus other outcomes). Team project. Method Eight EEG channels resampled to 100 Hz, 15-second segments every ten minutes, two LSTM layers (64 and 32 units) with dropout and early stopping. Assessment Scores too high to be credible (96.875% accuracy, recall of 0) because patients were not split across sets. An experimental prototype, presented as an approach; next step: a sound evaluation protocol. 
