Michail Berjaoui

Lead AI/LLM Engineer

5+ years shipping AI systems to production: enterprise RAG, LLM fine-tuning (LoRA/QLoRA), OCR/NER, and the MLOps to keep them reliable. Team lead (4-8 engineers). Based in Tokyo — open to international teams and French companies.

Enterprise RAG LoRA / QLoRA Fine-tuning vLLM / TensorRT-LLM Qdrant / Hybrid Search Kubernetes FastAPI GCP / AWS / Azure OCR / NER

Selected Work

70B LLM · Live Personality Switching
Fine-tuned a 70B model with dynamic LoRA adapters for an assistant that switches personality live in production.
2×H100 (Scaleway) · vLLM · LoRA hot-swap · Freelance, 2025
Multi-tenant Enterprise RAG + Text-to-SQL
Sovereign LLM deployment and multi-tenant RAG platform for enterprise clients, load-tested at 100 concurrent users.
Qdrant · Kubernetes · Scaleway · SOMA / Jenius
Identity Document Validation Pipeline
Automated ID validation combining OCR + NER + biometric matching, exposed as a production API.
FastAPI · Kubernetes · MinIO · SOMA / Rapid Visa
Geospatial ML on the National Fiber Network
Industrialized geospatial algorithms on Vertex AI — −40% uncertainty on France's national fiber network.
Vertex AI · BigQuery · Looker · Orange
E-commerce Semantic Search
Semantic search engine with RAG auto-activated beyond 100+ products and auto-scaling inference.
FastAPI · Qdrant · vLLM · Freelance, 2025
Financial Document Intelligence
Automated SCPI fund-report parsing (PDF → financial ratios), strategic clustering & recommendation.
PDF parsing · Clustering · SOMA / StoneLake
This site is part of the portfolio
MIRROR is not a template. It is a live, cost-engineered AI product: a RAG chatbot over my own knowledge base (Grok API + hybrid BM25 retrieval), document upload, web scraping, multi-user sessions — running on a small VPS for under $1/day. The architecture decisions are documented like I would for any client.

Technical Skills

Generative AI
Enterprise RAG LoRA / QLoRA ReAct Agents Text-to-SQL LLM Evaluation LlamaIndex LangChain
Inference & Scaling
vLLM TGI TensorRT-LLM AWQ / GPTQ Continuous Batching Paged Attention API-first (Grok, OpenAI-compatible)
Retrieval & Vector Search
Qdrant PGVector Hybrid Search (BM25 + dense) Reranking SQLite FTS5 Metadata Filtering
MLOps & GPU
Kubernetes Docker Terraform Helm Prometheus Grafana Scaleway H100 RunPod
Backend & Data
Python FastAPI / Flask BigQuery Spark Airflow MinIO OCR / NER PDF Parsing
Cloud
GCP (Vertex AI, GKE) AWS (Lambda) Azure OVH Scaleway Cloud Run

Experience

Freelance - Lead LLM Engineer
Remote
June 2025 - Present
  • Fine-tuning 70B model + dynamic LoRA adapters - assistant with live personality switching in production on 2×H100 at Scaleway
  • E-commerce semantic search engine - RAG auto-activated at 100+ products (FastAPI + Qdrant + vLLM + auto-scaling)
SOMA - Lead Data Scientist & AI Architect
Paris / Remote · Team Lead (4-8 people)
Jul 2023 - June 2025
  • Rapid Visa - Automated identity document validation: OCR + NER + biometric matching + FastAPI + Kubernetes + MinIO
  • Jenius - Sovereign LLM (Scaleway) + multi-tenant enterprise RAG (Qdrant) + Text-to-SQL + K8s + load testing 100 concurrent users
  • StoneLake - Automated SCPI report parsing (PDF → financial ratios), strategic clustering & recommendation
Orange - Data/AI Division
France
Jan 2022 - June 2023
  • Industrialization of geospatial algorithms on Vertex AI - 40% reduction in uncertainty on the national fiber network
  • BigQuery pipelines + Cloud Functions + Looker dashboards
Amadeus - Data Engineer
Nice / Sophia Antipolis
Sept 2020 - Oct 2021
  • Central Spark/Hadoop data lake + strategic KPIs for cloud costs & tool availability
Education
Université Paul Valéry - Montpellier
Master MIASHS - Applied Mathematics & Computer Science
2019 - 2021

Université Nice Sophia Antipolis
Licence Mathématiques Appliquées (avec mention)
2015 - 2018
Languages
French - Native
English - B2, full professional proficiency
Japanese - Learning
Certifications
How Google does Machine Learning
Coursera | Google - Jul 2022

Modernizing Data Lakes & Data Warehouses with GCP
Coursera | Google - Jul 2022