Développeur Freelance Full-Stack | IA & Produit

Conception et développement de zéro jusqu'à la mise en production d'une plateforme SaaS complète pour des cabinets d'avocats

• Conception d'une architecture multi-tenant avec isolation complète des données par cabinet.
• Mise en place de la gestion de dossiers, d'utilisateurs multiples et de droits par rôle (associé, collaborateur, secrétaire, stagiaire).
• Développement d'une fonctionnalité de génération documentaire sur modèles personnalisés pour 22 types d'actes et factures.
• Intégration d'un assistant juridique IA (JuristAI) s'appuyant sur 58 codes et 15 979 articles indexés.
• Implémentation d'un système de RAG avec refus de répondre en cas d'insuffisance.
• Automatisation de l'extraction de données depuis des documents PDF (parties, tribunal, dates).
• Configuration de notifications WhatsApp et email automatiques sur les décisions de justice.

GenAI Architecture and RAG Pipeline Design

• Architected production RAG systems end-to-end: adaptive chunking per content type, Google Embeddings via Vertex
• AI and OpenAI embeddings, vector database indexing with pgvector and FAISS, hybrid dense and sparse retrieval,
• semantic reranking, context budget management, and LLM-powered generation with structured output validation
• Designed and deployed multi-agent orchestration systems using LangChain, LangGraph, and MCP: defined agent
• roles, tool surfaces, state management, planning and reasoning loops, inter-agent communication protocols, and
• Human-in-the-Loop placement for high-stakes decisions
• Optimized LLM cost and latency through intelligent routing: deployed local models (DeepSeek 27B, Gemma, Llama)
• via llama.cpp and vLLM on private infrastructure with quantization, routing low-sensitivity tasks locally and high-
• complexity tasks to cloud APIs, reducing inference costs significantly
• Fine-tuned LLMs with LoRA and QLoRA on domain-specific datasets: dataset curation, supervised fine-tuning,
• systematic evaluation against benchmarks before and after, with production deployment and monitoring
• Evaluation, Safety and Explainability
• Built LLM evaluation frameworks using RAGAS for RAG pipeline assessment, DeepEval for pytest-style regression
• testing, and Giskard for vulnerability scanning covering prompt injection, PII leakage, and bias detection, alongside
• custom LLM-as-a-Judge pipelines with golden test sets, automated regression suites gating CI/CD deployments, and
• LangSmith tracing for end-to-end observability
• Designed explainability and traceability mechanisms for regulated environments: confidence scoring on all model
• outputs, structured audit trails logging every agent decision and tool call, HITL escalation paths for low-confidence
• predictions, and compliance documentation enabling full regulatory review of AI behavior
• Implemented safety guardrails: prompt injection detection, output filtering, PII redaction, allowlist/denylist policies, and
• blast-radius controls preventing high-risk autonomous agent actions in production
• LLMOps and Production Engineering
• Managed LLMOps lifecycle with ClearML and MLflow: prompt versioning correlated with evaluation scores, model
• versioning, A/B testing of agent strategies, drift detection, and automated alerting on performance degradation
• Deployed containerized AI services with Kubernetes, Docker, GitHub Actions CI/CD, and dev/staging/production
• environment management with quality gates. Cloud on AWS and GCP with Airflow orchestration, Kafka event
• streaming, and Grafana monitoring
• Built computer vision systems: YOLOv8 for real-time brand and logo detection across advertising images and video
• frames, with OCR pipelines for document intelligence and structured data extraction

Major Banking Institution | KYC AI in Regulated Finance

• Architected and delivered a production KYC AI system for a major financial institution: deep learning OCR pipeline for
• document field extraction, ML-based information matching against banking records, and computer vision signature
• authenticity verification, with full audit trails, confidence scoring, and HITL escalation for compliance in a high-stakes
• regulated environment
• Industrial and Pharmaceutical AI
• Built predictive maintenance ML models for Yazaki industrial operations on live sensor data streams: feature
• engineering, anomaly detection, failure forecasting, and FastAPI delivery of predictions to operations teams
• Developed deep learning classification models for pharmaceutical drug categorization and scoring, with rigorous
• evaluation frameworks and production deployment