MS Artificial Intelligence @ SJSU · Computer Vision · Agentic Systems · Edge AI
Probabilistic orbital risk framework over 13.59M VCM snapshots across 19,300 tracked objects. Transitions debris tracking from deterministic Keplerian propagation to uncertainty-aware 3D probability volumes. RK4 J2 propagator + TCN sequence model predicts residual trajectory errors rather than absolute coordinates, feeding a Dynamic Probabilistic Occupancy Map for launch window optimization. Pipelines in Polars; 41 engineered features across AMR, decay rates, Harris-Priester atmospheric density, and solar flux indices. Built and diagnosed three baselines before the physics-ML hybrid pivot — including a tautological RF classifier and a TCN with GAP signal dilution.
PyTorch Polars RK4 numerical integration TCNs irregular time series
Real-time PPE compliance and zone-based safety enforcement on SiMa.ai Modalix edge hardware (50 TOPS, 8× Cortex-A65). 10-class YOLOv8 detector with inverse-frequency class weighting to handle severe imbalance (person:vehicle ≈ 1:7.57). ByteTrack multi-object tracker enables per-track dwell-time measurement inside safety zones defined as camera-space polygons. ONNX export with static 640×640 input shape for MLSoC deployment. 18-run hyperparameter search (model size × LR × optimizer) tracked in W&B; best model: yolov8m, mAP@0.5 = 0.705.
YOLOv8 ByteTrack ONNX edge inference RTSP W&B
Asymmetric autoencoder for learned video compression targeting aerial surveillance. Shallow encoder (depthwise separable convs) runs on-device; heavy decoder (residual conv blocks) reconstructs server-side. Factorized entropy bottleneck with learned Laplacian priors for BPP estimation. GOP=10 with I-frame direct reconstruction and P-frame residual coding; temporal coherence loss stabilizes latent space across frames. Rate-distortion training: SSIM + L1 weighted with lambda-scheduled BPP. Trained on VIRAT aerial video.
PyTorch entropy bottleneck rate-distortion optimization SSIM depthwise separable convs
Agentic Perception Co-Pilot for Autonomous Vehicles — LangGraph ReAct agent reasoning over uncertain YOLOv8/DETR detections on nuScenes data. Advising: Prof. Kaikai Liu, SJSU CMPE.
- Machine Learning-Based Space Risk Management — IEEE ICEPES 2024 · ieeexplore.ieee.org/document/10653497
- Deep Learning Based Dementia Detection on MRI Data — Springer ICETSS 2024 · link.springer.com/chapter/10.1007/978-3-032-11488-4_15
Python · PyTorch · FastAPI · LangGraph · Docker · Postgres · Redis · YOLOv8 · ONNX · Polars · Airflow · dbt · DuckDB