Binh Phan

Developed a Python anomaly detection service using an LSTM autoencoder to score anomalies for sensor clusters across multiple locations along a river system in Indonesia. Ingests water level, flow rate, and wind readings over MQTT at 30-second intervals. Reduced a 45-minute cold-start delay by backfilling the most recent 90 data points at startup so the model can produce scores immediately. Built a worker pool for concurrent device streams and containerized the service with Docker for consistent deployment.

Built a FastAPI retrieval-augmented generation (RAG) platform for clinicians to query and summarize medical reference materials. Combined Weaviate vector search with a Neo4j knowledge graph to resolve relationships that vector similarity alone could miss. Developed a document ingestion pipeline for multi-modal extraction from large PDF sources (hundreds of pages), including a custom Vietnamese preprocessing stage to improve tokenization accuracy for Vietnamese text.

Built a Clinical RAG backend with FastAPI integrating Weaviate vector search and Neo4j knowledge graphs to support clinician question answering and summarisation. Developed a document ingestion pipeline with multi-modal extraction and custom Vietnamese preprocessing to improve tokenisation quality for long PDFs. Also created a flood detection service using a Python LSTM autoencoder, ingesting sensor readings via MQTT, scoring anomalies from multi-location clusters, and containerising the deployment with Docker and a worker pool for concurrent device streams. Reduced cold-start latency by backfilling recent historical readings so the service becomes productive immediately after deployment.