### BuitenBeter AI Engine

Built the BuitenBeter AI engine, a Python/FastAPI microservice using LangChain and Hugging Face for NLP and computer vision tasks.

- Added AI features to improve existing reporting workflows.
- Built a FastAPI-based REST API to serve machine learning, computer vision, and generative AI models for various citizen-report processing tasks.
- Built LangChain and LangGraph workflows to classify Dutch municipal issues into structured outputs.
- Deployed NLP pipelines with Hugging Face including mDeBERTa-v3 for zero-shot classification, named entity recognition, and sentiment analysis.
- Built a computer vision suite using YOLO, Florence-2, CLIP, and BLIP for reporter image classification, captioning, OCR, and visual question answering.
- Developed privacy-masking endpoints using deep neural networks, Haar cascades, and YOLO to detect and blur sensitive visual data such as faces and license plates in uploaded media.
- Added asynchronous inference support with a GPU-bound inference container, Azure Service Bus, and SignalR.

### AI-Powered Resume Matching Platform

An AI-powered platform that analyzes and matches resumes against job criteria using document processing and asynchronous AI workflows.

- Built an AI-driven matching platform with an Angular and Tailwind CSS frontend and a Python/FastAPI backend.
- Integrated Gemini AI to analyze and match resumes, using PyPDF to extract text from uploaded documents.
- Designed a decoupled microservices architecture with Azure Service Bus and standalone worker nodes to process heavy AI workloads asynchronously.
- Structured NoSQL data storage with Azure Cosmos DB and MongoDB, while using Azure Blob Storage for resume file management.
- Implemented SignalR to push real-time processing updates to the client, and used Chart.js and Mermaid to render matching metrics and workflow diagrams.
- Packaged the UI, API, and AI worker services into separate containers with Docker Compose, and used shell scripts to automate cloud deployment.