Landry Kotto

Context: As part of the launch of a super app combining mobile plans, financial services, and digital content (movies, games, news), I collaborated directly with Orange France's AI/Analytics team to develop and industrialize several machine learning solutions aimed at boosting user acquisition, retention, and engagement. Project 1: Targeting users likely to leave positive store reviews on GCP(Vertex AI); +20% increase in store ratings, leading to +7% in acquisitions - Collected and preprocessed multi-source historical app store data (BigQuery, HDFS, Oracle DW). - Engineered features and built a predictive clustering model with drift monitoring. - Automated weekly training and daily prediction pipelines on GCP VertexAI. - Integrated predictions with the campaign management tool to deliver targeted user campaigns (SMS, email, push, in-app). Project 2: Development of an internal chatbot for the Product team and Data Analysts, centralizing multi-source customer feedback (store reviews, forms, transcribed hotline calls) to accelerate feature prioritization and quickly identify root causes of performance drops - Ingestion & preprocessing of textual data (cleaning, normalization). - Feedback chunking for fine-grained search granularity. - Vectorization using Sentence-Transformers / DeBERTa-v3 and storage in PgVector / FAISS. Semantic search (retriever) to quickly surface relevant feedback. LLM-powered generation (GPT/LLaMA) to deliver contextualized and actionable summaries. Project 3: KPI monitoring through interactive dashboards – Real-time tracking - Built dynamic dashboards using Google Looker Studio to track strategic KPIs generated by the super app. - Enabled real-time updates and a self-service interface to support data-driven decisions by business teams in France and Africa. I also supported the team on other projects by building custom diversification methods on Dataiku DSS, which reduced churn by 5 points, and by conducting performance drop analyses.

Context: Eneo, a subsidiary of the British group Actis, is a leading company in the production and distribution of electrical energy. At the beginning of each month, technical agents go on-site to manually record the electricity consumption index in each household. The customer's bill is then calculated based on the current and previous index readings. However, this manual process results in a high complaint rate — estimated at 35% each month — which is unacceptable for the company. Project 1: Lead on the development of a deep learning–based OCR system to improve billing accuracy; this led to reduce complaint rate from 35% to 10%. - Imported and preprocessed a dataset of several hundred meter images collected from field operations (OpenCV: deduplication, noise reduction, filtering). - Labeled key elements in the images (brand, meter number, manufacturing year, consumption index) using PPOCRLabel (PaddlePaddle tool) with bounding boxes. - Researched and implemented machine learning algorithms to automatically detect text regions in images. - Extracted and read textual information using OCR techniques. - Classified extracted zones based on their content. - Built a Qt-based GUI to enable the billing team to visualize and use the extracted data. Project 2: Development of billing performance KPIs with Power BI. - Designed key performance indicators related to data completeness, billing accuracy, and customer satisfaction.

Context: Developed NLP/LLM models for a U.S.-based educational platform to automatically analyze student answers in natural language. The system supports ~50,000 students across 432 cohorts and ~100 questions, enabling facilitators to monitor answer sharing and automate correction. Weekly training pipelines were implemented via Azure Functions. Project 1: Detection of Similar Answers per Question Identified students sharing answers for each question and generated alerts for facilitators. - Detected similarity between answers using cosine distance to generate alerts. - Automated the weekly training pipeline via Azure Functions, integrating new answers and improving accuracy. - Stored embeddings in Parquet format, optimizing matrix size (~1,000 responses × 300 dimensions per question) and ensuring system scalability. - Deployed a prediction pipeline via an Azure endpoint, taking question ID and text answer as input to check similarity in real time. Project 2: Classification of Correct / Incorrect Answers (Pass/Fail) Automatically classified student answers as correct (pass) or incorrect (fail) to enable correction. - Preprocessed and normalized student answers in natural language. - Developed an NLP/LLM classification model to determine pass/fail per question. - Evaluated model performance (precision, recall, F1-score) to ensure reliability. - Automated the weekly training pipeline via Azure Functions to maintain an up-to-date model. - Deployed an Azure prediction endpoint, taking question ID and text answer as input to return the classification in real time.