KhaifParis · Full remoteENFR

Si-ThamiKhaif

GenAI Engineer / LLM Systems / MLOps

I deploy LLM agents on your own infrastructure.Data sovereignty without the efficiency hit.

01about.me

Auto-mate AgencyFounder and GenAI Engineer
2023 to today · full-time since January 2026 · Paris, remote
- Generative AI systems for clients, from scoping through to handover to their own teams.
- RAG assistants over client document bases; automated extraction and summarisation of contracts, invoices and reports.
- Advisory on safe adoption and on-premise LLM deployment, for teams whose data cannot leave the network.
- Typical result: 40 hours of manual work removed per month.
Python · LangChain · LangGraph · OpenAI · Mistral · Anthropic · vector databases
Air Liquide R&D JapanMachine Learning Engineer, R&D
June to December 2025 · Tokyo
- Design, build and evaluation of a production RAG system, with hallucination rates measured and LLM-as-a-judge validation before go-live.
- Rebuilt a MedTech facial measurement pipeline, from 2D photo to real-time 3D video, 468 landmarks per frame, 95% accuracy.
- Owned the data foundation of a global ML project: scraping and preprocessing pipelines under CI/CD.
LangChain · FAISS · Pinecone · LLM-as-a-judge · Python · MediaPipe · GitHub Actions
Safran AerosystemsSoftware AI Developer, internship
June to September 2024 · Plaisir
- Fully on-premise RAG assistant deployed on the company SharePoint. 250+ active users, around two hours saved per person per week.
- Delivery tracking dashboard and delay forecasting models for the supply chain team.
LangChain · Ollama · SharePoint · Power BI · Power Query · VBA
Safran Aircraft EnginesSoftware and Data Automation, internship
June to September 2022
- Python automation and a delivery tracking tool shared across teams.
Python · Pandas · Scikit-learn

02my-projects

01Local vs cloud RAG (benchmark)
code
Question answering over a document corpus: every claim cites the passage backing it, and the system declines to answer rather than invent one. Nothing leaves the machine by default; a hosted backend puts a number on the local/cloud gap.
Python · Ollama · Azure OpenAI · FastAPI · NumPy · rank-bm25 · sentence-transformers · Docker
02LLM credit scoring
code
Business loan applications arrive as free text. A local LLM extracts the figures, then plain Python scores them on an eight-criterion grid — the model never does the arithmetic. 97.3% field accuracy, no changed recommendation over 90 runs.
Python · Ollama · Llama 3.1 8B · Streamlit · pytest
03Cyber AI NER
code
Named-entity extraction from English threat reports: 4,876 annotated sentences, a three-type BIO scheme. Two approaches compared, a bidirectional LSTM built from scratch against a fine-tuned pre-trained transformer. Done as a pair.
Python · TensorFlow · Keras · simpletransformers · BERT · scikit-learn · Parquet