Experience
Technical ownership, production delivery, research, and teaching
ML Engineer
Tensor Auto · San Jose, CA
Sole technical owner of three ML/CV systems spanning sensor calibration and occupant perception. I independently design and train the models, build their data and training pipelines, and integrate onboard inference with infrastructure and testing teams. Deployed two calibration systems to production and completed real-vehicle evaluation of integrated occupant perception.
Product Definition & System Design
Partnered with my manager to translate high-level product direction into concrete system requirements, evaluation criteria, integration contracts, and delivery milestones.
Neural Network Design & Training
Designed and trained camera–LiDAR calibration with camera-conditioned FiLM and cross-modal attention, reducing rotation error by 95% (2.95° → 0.14°) on an unseen vehicle. Built LiDAR–LiDAR calibration from scratch with shared spherical encoders and circular azimuth correlation, reducing mean rotation error by 83% (4.86° → 0.83°) in testing.
Automated Sensor-Data Infrastructure
Built and launched a shared Airflow/Kubernetes collection pipeline across two regions, streaming 17+ TB of raw drive logs through single-pass extraction to produce ~600 GB of camera–LiDAR and LiDAR–LiDAR calibration data. Added drift-aware synchronization, traceable calibration labels, and per-log retries; reduced the runtime image from 20 GB to 600 MB.
Production Deployment
Deployed both calibration models through PyTorch → ONNX → TensorRT → C++, with numerical-parity checks, robust temporal aggregation, and sensor-rig updates. Fleet evaluation showed 44% / 57% lower rotation/translation error for camera–LiDAR calibration and 63% (0.93° → 0.35°) lower rotation error for LiDAR–LiDAR calibration.
Detector Adaptation & Evaluation
Fine-tuned RTMO-s for front- and rear-seat person detection and keypoints using targeted cabin annotations mixed with COCO data. Rear-seat forward-lean confirmation increased from 72.1% to 98.6% on reserved evaluation sessions, without increasing empty-seat false positives.
Integrated Occupant Perception
Designed front- and rear-seat occupant monitoring with a ResNet-18 multi-task occupancy/posture model, RTMO detection, iToF reasoning, and VLM temporal fusion. Connected confidence-aware outputs to the vehicle interface; system benchmarks measured 99.3% occupancy F1 and 95.8% posture Macro-F1.
Technical details and results are documented in the corresponding case studies.
Research Assistant
University of Rochester · Jan 2023 — May 2024 · Advisor: Professor Tong Geng
- Developed PyTorch-based graph learning and nature-inspired computing methods, combining algorithm design, mathematical analysis, and controlled experimentation across multiple application domains.
- Designed reproducible research workflows for model comparison, performance analysis, and collaborative experimentation in shared Linux computing environments.
- Led a five-member research infrastructure effort covering Linux server administration, automation, Docker/Kubernetes environments, and reliable access to shared compute resources.
Teaching Assistant
University of Rochester · 2022 — 2024
Delivered lectures, discussion sessions, office hours, and laboratory support across multiprocessor systems, memory architecture, computer organization, and electronic circuits. Guided students through technical work in C, MIPS assembly, Verilog, SPICE/LTspice, and hardware debugging.