Final-year engineering student at ESSAI (Tunis), interested in applied AI, NLP and LLM systems. I care as much about knowing when a system can't be trusted as about making it work.
- Evaluation & reliability: Measuring whether a system actually works, and where it fails. Fine-tuning NLLB for English to Arabic showed me a model can improve in-domain and get worse out-of-domain, and you only see it if you test on both.
- LLMs & NLP: Applications around language models, retrieval and multilingual text.
- Cost-aware AI: When is a cheaper model enough? I'm particularly interested in the challenge of deciding when a smaller model is sufficient, and how to evaluate that decision.
- Looking for a 4–6 month PFE starting between January and March 2027, in applied AI, NLP or data.
- I try to document evaluation scope, design choices and known limitations in my projects
- Languages: Arabic (native), English (good), French (intermediate).
- The mathematics behind machine learning: linear algebra, gradients and probability.
- Transformers and LLMs from the theoretical side.
- Linux and SQL.
Programming: Python Machine Learning: scikit-learn, XGBoost, SHAP, PyTorch LLM & NLP: Hugging Face Transformers, RAG (BM25 + dense retrieval + reranking), Groq API, ChromaDB, sacreBLEU Document AI: PaddleOCR, EasyOCR Tools: Git, Jupyter, pytest, Gradio
| Project | What it does | Stack |
|---|---|---|
| tunisian-labor-law-rag | French-language assistant over the Tunisian Code du Travail and CNSS FAQ; a verifier declines unsupported answers. Small evaluation (21 gold questions) | Python · BM25 · bge-m3 · ChromaDB · Groq |
| massive-routing | Routes between a small and a large LLM for intent classification. 77.3% vs 78.0% for the large model alone on 150 test requests, at lower estimated cost | Python · scikit-learn · Groq |
| receipt-intelligence | Receipt photo to structured record, with arithmetic checks that flag records it can't trust. Tested on 500 CORD receipts | Python · PaddleOCR · LLM extraction |
| nlp-en-ar-translation | Fine-tuning NLLB-200 for English to Arabic. One run: better in-domain, worse out-of-domain | PyTorch · Hugging Face |
| diabetes-readmission | Predicting 30-day readmission on 101,766 hospital encounters; four tuned models compared (best ROC-AUC 0.69) | scikit-learn · XGBoost · SHAP |
📫 Contact: khadija.trabelsi.003@gmail.com