Richard Tang

• Built an agentic AI system that synthesizes web, operational, and internal-knowledge sources into operations-optimization recommendations.
• Deployed it as a streaming OpenAI-compatible FastAPI service on AWS ECS Fargate behind an Application Load Balancer, with CPU and request-based target-tracking autoscaling.
• Designed real-time streaming anomaly detection over flowrate, temperature, and humidity sensor data feeding internal operational metrics.
• Built a synthetic-data pipeline that injects a configurable fault taxonomy into clean time-series, producing synthetic labeled datasets to train and validate long-horizon trend-detection models efficiently.

Python · LangGraph · FastAPI · AWS ECS Fargate · ALB · Time-Series

• Engineered a rocket-detection framework using YOLOv26 and classical computer vision, achieving 84% mAP on internal benchmarks at 30+ frames per second.
• Deployed the detection framework on a Jetson Nano with an active gimbal for real-time rocket tracking.
• Built a Webots simulation environment to generate large-scale training and evaluation data for object-detection models used in drone deployment.

YOLO · OpenCV · Jetson Nano · Webots · Object Detection

• Combined geometric 3D encoders (CUT3R, VGGT) and fusion strategies to improve egocentric video spatial reasoning in a Qwen vision-language model, with supervised fine-tuning.
• Contributed to a Blender-based simulator that renders photorealistic egocentric household-task video to generate causal spatial-reasoning QA for evaluating multimodal models.

PyTorch · Vision-Language Models · Spatial Reasoning · LlamaFactory · Blender

Combined deterministic ARC pattern solvers with an agentic harness where LLM agents propose minimal ONNX graph specs through tool use, gated by byte-exact verification with escalating models and synthesis handoffs between attempts.

Post-trained a Nemotron-3-Nano-30B-A3B model with supervised fine-tuning and GRPO for NVIDIA's logical-reasoning benchmark, and built deterministic solvers that generate synthetic, verified chain-of-thought traces as training data.