Yvan Kondjo

Multimodal RAG AI Agents for 24/7 Customer Support on Instagram and WhatsApp - Developed RAG-based conversational AI agents automating 24/7 customer support on Instagram and WhatsApp, leveraging a knowledge base for accurate answers and performance tracking via LangSmith. - Designed and deployed a full-stack application with a FastAPI backend and Next.js frontend, hosted on Google Cloud Run and Vercel, using Supabase, Redis, and pg_vector for efficient data management.

AI & Data-Driven Optimization – Renault After-Sales Services ➡ Development of a hybrid search engine for technical bulletins - Designed and deployed a hybrid search engine combining BM25 and dense embeddings (TextGecko) to automatically extract solutions from customer feedback. - Deployed on MongoDB, achieving a +60% improvement in search performance. ➡ Multilingual architecture for the DAB (Diagnostic Assisté du Bruit) application - Built a multilingual architecture translating the French corpus into 27 languages with a per-language index to optimize latency and relevance. - Developed an internal web app to benchmark translation models and created a multilingual glossary from 27,000 domain-specific documents. ➡ Automation of customer complaint categorization - Developed a classification pipeline using Gemini 1.5 Pro to automatically tag customer complaints by issue type. - Processed 1,000 complaints with 88% precision, 90% recall, and a 200% reduction in manual analysis time. ➡ ETL for analyzing mechanics’ behaviors on after-sales applications - Designed a BigQuery data warehouse aggregating over 100,000 logs from 4 Renault after-sales applications. - Built a Python/SQL data pipeline for real-time monitoring and Spotfire visualization to uncover key insights into app usage patterns.

Developed an AI system to automate the quality control of electronic PCB boards in an industrial production environment. - Fine-tuned a YOLOv8 pre-trained model on Cloud GPU (A10) for automated defect and missing component detection, using LabelImg annotation and data augmentation to improve model robustness. - Built and deployed a REST API (FastAPI) integrating the fine-tuned model into the real-time quality inspection pipeline, achieving 80% accuracy and an F1-score of 60% in production.

Analyzed operational data to calculate KPIs such as the average number of packages received per city and region, enabling performance ranking of CAMPOST agencies nationwide.