Projects
Synthetic Animal Motion Generation
Developed an end-to-end system that reconstructs realistic 3D quadruped motion from monocular video, addressing the lack of scalable motion datasets for simulation and radar-based learning. The pipeline converts raw video into structured joint trajectories by combining segmentation, skeleton extraction, and SMAL model fitting, enabling consistent motion reconstruction across time.
The system processes video frames to extract silhouettes and structural representations, followed by keypoint estimation and parametric model fitting. The resulting motion is exported into a Unity-based animation system, where it can be replayed, controlled, and extended. This creates a full pipeline that bridges perception, modeling, and simulation.
This approach enables scalable generation of synthetic motion datasets with higher diversity and controllability than real-world capture, making it valuable for training machine learning systems that depend on structured motion data.
Tech: Python, PyTorch, OpenCV, Unity, SMAL
Radiation Signal Classification
Built a real-time machine learning system for classifying gamma-ray waveforms using a convolutional neural network with uncertainty calibration. The system improves signal detection in noisy environments and provides reliable predictions even in ambiguous scenarios.
The model learns temporal patterns from waveform data and incorporates Bayesian uncertainty estimation to improve robustness. The pipeline is optimized for low-latency inference, enabling continuous real-time processing of streaming signals.
Integrated with Unity and deployed to HoloLens, the system visualizes radiation signals in an interactive 3D environment, allowing users to intuitively explore signal patterns and spatial distributions.
Tech: PyTorch, Unity, HoloLens, C#
AI Portfolio Knowledge Base
Designed a retrieval-augmented generation system that transforms static documents into a structured, queryable knowledge base. The system enables semantic search and context-aware response generation for technical and behavioral queries.
Built a document ingestion pipeline with structured chunking and embedding-based retrieval using ChromaDB. Combined retrieval with LLM-based generation to produce structured and coherent responses.
Optimized prompt design and retrieval strategies to improve response consistency and reduce hallucination, enabling reliable information access for real-world use cases such as interview preparation.
Tech: LlamaIndex, ChromaDB, Transformers
PrairieLearn Feedback Engine
Developed an automated feedback system for proof-block ordering problems, enabling structured and scalable evaluation of logical reasoning in large classrooms.
The system analyzes student submissions, identifies logical ordering errors, and generates deterministic feedback based on rule-based evaluation. It provides meaningful guidance rather than simple correctness checks.
Deployed to 800+ students, the system significantly reduced grading workload while improving learning outcomes through immediate and structured feedback.
Tech: Python, Mustache.js