Hassine E.

I build LLM pipelines that turn PDFs, web pages and messy files into clean, structured data, and RAG assistants that answer questions from your documents, with sources. As founding ML engineer at an Italian startup, I built the LLM-agent engine that extracted curricula from 50+ European universities into one JSON schema at 95% data precision. WHAT I CAN DO FOR YOU - Extract structured data (JSON, CSV, SQL) from PDFs, HTML and scraped pages with LLMs, including validation rules and a measured accuracy report - Build RAG chatbots and semantic search over your documents (LangChain, vector databases, OpenAI/Anthropic APIs or open-source models) - Build LLM agents and automated ETL pipelines in Python (Scrapy, Playwright, PySpark, SQL) - Fine-tune open models with LoRA/QLoRA (Hugging Face, PyTorch), including vision-language models - Train and evaluate ML and time-series forecasting models (scikit-learn, XGBoost, Optuna) RECENT WORK - Startup (Founding ML Engineer): LLM-agent extraction engine, RAG and automated ETL pipelines over multi-format university data. The platform reached about 1,000 sign-ups in its first 2 days. - MSc thesis (Fondazione LINKS): built a LLaVA-style vision-language model (CLIP ViT-L/14 + LLaMA-2-7B, QLoRA) on a single 11 GB GPU. Reducing the image-text "modality gap" improved every metric: +9% CLIPScore, 16% fewer hallucinated objects, +13% object recall. - R&D Data Scientist (Enerbrain): end-to-end digital twin for building HVAC control, covering system identification, a Gymnasium simulator, MPC and Soft Actor-Critic RL (RLlib, Optuna, AWS). Validated on 5 real buildings. - Teaching Assistant, Politecnico di Torino: Python and Java labs for 300+ students. I can explain technical work clearly to non-specialists. HOW I WORK 1. We agree on the goal, the data and what "good" means (e.g. which fields, what accuracy). 2. First milestone is small and measurable, for example extraction on 20 of your documents with an accuracy score, before we scale. 3. You get clean, documented Python in your repository, with a README and a short handover call. 4. I send updates every 1-2 working days and flag problems early, not at the deadline. I work with your API keys and infrastructure and keep your data confidential. Tools: Python, PyTorch, Hugging Face (Transformers, PEFT), LangChain, scikit-learn, XGBoost, RLlib, PySpark, SQL, Scrapy, Playwright, Docker, Git, Linux, AWS. Education: MSc Data Science & Engineering and BEng Computer Engineering, Politecnico di Torino. Before moving into AI, I completed 17 graphic-design jobs on Upwork (2020-2023), almost all rated 5.0. The reviews below come from that work, and I bring the same reliability to AI projects. Based in Turin, Italy (CET). English (C2) and Arabic (native); French and Italian at an intermediate level. Send me a few sample documents and the fields or answers you need. Within one working day, I'll reply with how I would approach it and a fixed first milestone.