Melek

AI Engineer

PwC France

Apr 2026 - Jun 2026 • 2 mos

I designed and built a multi-agent AI platform for automated HR document generation and enterprise knowledge retrieval. The system orchestrated autonomous workflows (LangGraph) for HR article generation, document analysis, and legal reasoning. The retrieval layer used PGVector with semantic and hybrid search over several thousand internal HR documents and regulatory references, combined with Azure OpenAI, Human-in-the-Loop review, and LLM-as-a-Judge evaluation to keep generated content accurate. I also built the engineering layer: streaming/async FastAPI APIs, an asynchronous document-ingestion pipeline via RabbitMQ, and OCR/vision-based semantic chunking. The platform was deployed with Langfuse observability and Docker/Helm/Kubernetes so the client's own team could run it after handoff.

Data Scientist / AI Engineer

Crédit Agricole

Apr 2025 - Jan 2026 • 9 mos

I worked on a climate-risk scoring model assessing physical and transition climate risks for the group's corporate clients. I collaborated directly with the data science team on designing and improving the climate & environmental (C&E) scoring models, then built the engineering around them: high-performance REST APIs in FastAPI exposing the calculation and prediction engines, with asynchronous handling and automatic documentation. I industrialized the solution under MLOps principles — reproducible environments and automated deployment instead of manual releases — containerized every service with Docker for consistency across dev/test/production, and automated Kubernetes deployment with ArgoCD.

Data Scientist

Imerys

Apr 2024 - Nov 2024 • 7 mos

As a Data Scientist at Groupe Imerys, I built an intelligent quality-control and predictive-maintenance system for industrial machines based on time series and computer vision. I designed a computer-vision system (YOLOv8, Faster R-CNN) to classify and automatically detect anomalies on small, fast-moving parts, and built a Python framework to statistically evaluate model performance — precision, recall, F1, mAP — so improvements were measured, not assumed. On the predictive-maintenance side, I used industrial time-series and sensor anomaly detection (One-Class SVM, Elliptic Envelope, Prophet), and separately designed an LLM-based instructive chatbot for the production line. The whole system was deployed and industrialized on AWS, with live Grafana dashboards that line operators actually used.