Otmane El Bourki

• Built and productionized an AI references-discovery platform in 3 months (React, FastAPI, PostgreSQL/pgvector, Azure) for ALTEN’s sales teams and directors; 101 active users and 1,103 searches in early rollout. • Built the RAG pipeline (Mistral OCR and embeddings, OpenAI fallback) with an LLM reranker that selects the best client references and explains each choice; 80% R@1, 90% R@5, 0.85 MRR on a 100-question internal evaluation. • Built the platform’s evaluation harness (retrieval metrics, LLM-reranker checks) and championed eval-driven development across A3, surfacing the need for AI governance to the Chief AI Office.

– Delivered a client-facing multimodal document processing system leveraging OCR, vision-language models, and hybrid pipelines - automating invoice and receipt extraction, eliminating manual data entry, and enabling high-throughput, auditable workflows for finance and operations teams. – Architected a real estate platform (FastAPI, LangGraph, DeepAgents) orchestrating 6 specialized LLM subagents across property search, financial analysis, location intelligence, and market reporting - eliminating manual coordination across all research domains. – Built an open-source sandboxed runtime (codibox) for LLM-generated Python, implementing Docker-level isolation (namespace separation, read-only filesystem, resource caps) to eliminate unsafe execution risks and reliably capture output artifacts in agentic pipelines. – Implemented an enterprise GraphRAG platform for financial documents, integrating Neo4j knowledge graphs, hybrid vector/full-text retrieval, and LLM-based reasoning - providing clients with accurate, explainable answers for due diligence, compliance, and risk analysis.

– Architected and deployed a production-grade multi-agent AI platform using LangGraph and FastAPI, orchestrating six specialized agents (RAG, SQL, Deep Research, Data Analyst) — accelerating financial report generation to 30 minutes from half a day. – Designed and implemented an adaptive RAG pipeline with semantic chunking and hallucination grading, integrating Qdrant embeddings and multi-step retrieval validation to deliver high-accuracy document intelligence across hundreds of PDFs and Tagetik financial datasets. – Built a scalable multi-LLM orchestration layer supporting OpenAI, Anthropic, Gemini, Groq, and AWS Bedrock implementing supervisor-based routing, checkpointed state management, OAuth-secured APIs, and Dockerized microservices with CI/CD, designed with the analyst team to ensure reliability and auditability.

– Architected and deployed a production-grade multimodal RAG platform on GCP, fine-tuning LLaMA for domain-specific hotel guest intent detection and structured response generation; built scalable inference endpoints on Vertex AI and containerized APIs on Cloud Run. – Engineered a high-performance LLM serving stack using AWQ/GPTQ quantization and vLLM-based dynamic batching - increasing inference throughput 3x, reducing latency under load, and cutting infrastructure costs by 47% through autoscaling and GPU optimization. – Designed an enterprise-grade document ingestion pipeline supporting large, multi-format files, implementing schema-enforced structured extraction with Pydantic and automated embedding workflows to power hybrid (vector + lexical) retrieval with reranking. – Implemented end-to-end LLM observability, evaluation, and security hardening - including RAG quality metrics (faithfulness, relevance), P50/P95 latency monitoring, prompt-injection mitigation, and safe model rollouts via traffic splitting and A/B testing.

– Developed a Medical RAG QA chatbot using Meditron 7B, Qdrant vector database, and PubMedBERT embeddings - containerized, GPU-enabled, and deployed via Chainlit for evidence-grounded, real-time clinical Q&A over medical literature. – Built North, a dual-stack North African dialect classifier (scikit-learn + fine-tuned BERT) across Arabic and Arabizi Latin script, achieving 94.5% test accuracy. – Benchmarked SOTA CV architectures on FER2013, achieving 90% avg. precision and 20% generalization improvement via targeted augmentation and regularization.