Augustin Combes

Implementing end-to-end AI-powered solutions for various Galadrim clients, covering backend, frontend, and MLOps aspects. Current projects include: - Okantis: Served as AI strategist within a 5-person team building an intelligent medical data warehouse. Responsible for designing systems to automatically structure PDF documents into a database using online learning and implementing natural language querying for structured output generation (RAG-SQL). - Inventozen: Created image search capabilities for identifying abstract artworks. - Splitter: Built a predictive model forecasting 5-year real estate price trends across France. - MdB Multimedia: Delivered a full-stack click-and-collect and delivery platform for caterers. - Expertise France: Developed a RAG-powered chatbot connecting project owners with geographical data experts, capable of retrieving information from 150+ documentation files. - Garuda House: Created a self-reflective chatbot for theological Buddhist inquiries that retrieves and quotes information from 3GB+ of multilingual sources (Tibetan-French-English) using function calling capabilities.

- Acquired and curated data periodically from over 150 e-commerce clothing websites, ensuring high-quality information. - Integrated scraped data continuously into AWS services (RDS and S3) and developed monitoring systems to ensure data quality. - Created organized Notion ticket systems to facilitate efficient collaboration among scraping developers.

Representation learning for x-ray radiology studies (sets of images appaired with expert free-text). Enhanced ViT-MAE unsupervised pretraining baseline on CLIP-like feature extraction evaluation. - Modeled radiology studies as a cohesive variable-length image set, efficiently allowing the modeling of intra-study inter-image interactions throught bottleneck tokens. - Enriched pretraining with a masked diffusion objective, incentizing reconstruction of images details, ultimately leading to extracted features having high predictive performance using linear probing.

- Data Privacy: Applied advanced Deep Learning techniques, specifically LSTM-GAN, to anonymize medical records. This allowed for the secure sharing of synthetic data that mirrors the distribution of the European Hospital Georges Pompidou's sudden cardiac arrest EHR database. - Predictive Fairness: Utilized the generative model to balance the database, enhancing the performance of the preventive detection model, especially for previously underrepresented minority classes.

Missions for NGOs and European development agencies: - Led a causal study (difference in differences) quantifying the impact of EU-sponsored canal renovation in Iraq on the agricultural production. Used Planet satellite images and ground-truth field labels as explanatory data for the econometric study. - Involved in a brand image study in Arabic countries. Creation of the database by scraping on Facebook and Twitter, NLP modelling (sentiment analysis / hate speech detection) using pretrained AraBERT LLM .

Development of a screen printing algorithm using machine learning methods allowing the preview of customized products on e-commerce websites.