Selected Work

Production systems, research, and selected technical projects.

As sole technical owner, I take production ML and robotics systems from initial requirements and data contracts through independent design, implementation, and validation—and, for deployed systems, onboard integration. The work below emphasizes measurable recovery, cross-domain robustness, and complete engineering ownership.

Industry Systems

PRODUCTION CALIBRATION PORTFOLIO

Neural Sensor Calibration Systems

Two independently developed and deployed systems covering camera–LiDAR and LiDAR–LiDAR extrinsic calibration—recovering large rotational disturbances to sub-degree residuals, scaling training through a reusable distributed runtime, and carrying the same geometry into production C++.

2production systems deployed
500+ GBmultimodal sensor data
10° → <1°controlled recovery
Scale-outreusable distributed training runtime
01 · LCCNET · CAMERA ↔ LIDAR
Production · Deployed

Neural Camera–LiDAR Calibration

A production 6-DoF calibration system that learns corrective geometry from RGB and LiDAR evidence, scales through a world-size-aware training runtime, recovers perturbations as large as 10° to sub-degree residuals, and closes the loop onboard.

500+ GBsensor data
50K+curated pairs
0.31°fleet validation
<0.6°held-out vehicle
Synthetic LCCNet comparison showing predicted calibration, a visibly rotated mis-calibrated input, and reference alignment
QUALITATIVE RECOVERYFully synthetic · click to enlarge
02 · LILINET · LIDAR ↔ LIDAR
Production · Deployed

Neural LiDAR–LiDAR Calibration

A geometry-aware system that combines full-azimuth correlation with a production-hardened distributed training runtime, recovers cross-LiDAR alignment from large rotational disturbances, and ships through TensorRT/C++.

5K+curated pairs
+62.5%improvement
0.35°session held-out
95.9%frames improved
Synthetic LiLiNet comparison showing aligned main and blind-spot LiDAR clouds, a rigidly mis-calibrated input, and reference alignment
CROSS-LIDAR ALIGNMENTFully synthetic · click to enlarge
03 · IN-CABIN · OCCUPANT PERCEPTION
Active Development

In-Cabin Occupant State Perception

A dual-head perception system that jointly classifies seat occupancy and occupant posture from fisheye cabin imagery, with leakage-resistant data construction, deterministic evaluation, and a deployable ONNX interface.

01 · MODELSeat-aware multi-task inferenceshared visual representation for occupancy and posture classification
02 · DATASession-isolated data contractburst-aware sampling with recording-level separation
03 · VALIDATIONSafety-oriented error taxonomydirectional errors, per-class recall, and degeneration checks
04 · SYSTEMDeterministic deployment interfaceversioned preprocessing, tensor semantics, and output gating
Schematic architecture for per-seat occupancy and occupant-posture classification from fisheye imagery
CURRENT VISUAL BASELINESchematic · click to enlarge

Research & Selected Projects

Fast Robot Navigation & Control
Fast Robot Navigation & Control
Embedded Systems & Robotics

Real-time navigation, control, and obstacle avoidance on a resource-constrained embedded robot.

Point Forecasts to Probability Clouds — Probabilistic Electricity Price Forecasting
Point Forecasts to Probability Clouds — Probabilistic Electricity Price Forecasting
Research Project

Generative probabilistic forecasting of day-ahead electricity prices with calibrated uncertainty

Robotic Tool Handler
Robotic Tool Handler
Technical Project

Senior Design Project - ARM Cortex-based robotic system with real-time control

Robot Mapping, Estimation, and Interaction
Robot Mapping, Estimation, and Interaction
Technical Project

Autonomous navigation system with SLAM capabilities for TurtleBot2

LRHPerception - Monocular Real-time Perception for Autonomous Driving
LRHPerception - Monocular Real-time Perception for Autonomous Driving
Research Project

Unified perception pipeline achieving 29 FPS with object tracking, trajectory prediction, road segmentation, and depth estimation

Spatiotemporal-Linear - Universal Multivariate Time Series Forecasting
Spatiotemporal-Linear - Universal Multivariate Time Series Forecasting
Research Project

Novel forecasting model using Residual Neural Networks with Spatial Attention