Richard Chiou

Fine-tuned computer vision and deep learning models (ResNet-50) using PyTorch to identify species, individual animals, sex, and age-class from wildlife footprint images, achieving over 95% accuracy. Integrated optimized ONNX models into WildTrackAI mobile app, reducing prediction time from 2s to 300ms, enabling 1000+ field users. Deployed a BioBERT-based LLM model using LangChain-powered RAG on SageMaker to handle 500+ patient conversations per month, delivering personalized healthcare advice. Improved clinical workflow efficiency by presummarizing patient information from virtual assistant conversations using a BART-based Transformer, reducing average doctor-patient meeting time by 30%. Led a cross-functional AI team of 7, implementing reproducible MLOps with MLflow, DVC, Docker, and AWS S3, cutting deployment time by 40%.

Architected spatio-temporal forecasting models using Graph WaveNet to predict traffic flow one hour ahead, achieving 23% higher accuracy than ARIMA. Applied transformer-based temporal encoders for multi-horizon congestion prediction across heterogeneous inputs, enabling early evacuation bottleneck detection and reducing prediction error by 18%. Created real-time incident detection models combining loop detector data, GPS traces, and DOT feeds to classify crashes and stalled vehicles with 91% precision and 87% recall, reducing operator detection latency by 35%. Built reinforcement learning model for adaptive traffic signal timing and contraflow recommendations, validated in SUMO simulations with up to 18% improvement in throughput during peak evacuation load. Established real-time traffic AI pipelines on AWS, ingesting 500K+ daily GPS and sensor records through Kafka, preprocessing with Spark, and deploying models on SageMaker with 99.9% uptime for high availability.

Mentored learners in feature engineering, model evaluation, and deployment, resulting in measurable skill improvements and career placements for 70+ students in data science roles. Authored 50+ hands-on exercises and guided 20+ capstone projects, producing production-ready ML models and analytics solutions.