02 / Experience

From warehouse fleets to humanoids to autonomous vehicles.

  1. Pushing onboard VLM/VLA models up the curve with data: finding what moves the metrics, proving it with ablations, and letting scaling laws decide what comes next.

    • Drive performance improvements for onboard VLM/VLA models by closing the loop between data and evaluation: mine, label, train, ablate, repeat.
    • Use data-scaling trends to decide where the next gains come from: which scenarios to mine, what to label, and how much of each to add to the mix.
    • Mine rare and safety-critical scenarios from large-scale fleet logs with several complementary approaches, from heuristic pipelines to VLM-generated (Cosmos-3) captions and scene metadata.
    • Build agentic labeling workflows that turn mined clips into training-ready supervision.
    • Train models across ablations and data mixes for mid-training and fine-tuning, defining the sampling strategies, annotation schemas and quality filters behind them.
    Onboard VLM / VLAData scalingData mixes & ablationsAgentic labelingData miningCosmos
  2. Led data infrastructure and platform work for robot learning: the systems that take multimodal robot logs to training-ready data for foundation-model research on Atlas and Spot.

    • Unified RGB-D video, proprioception and force/torque streams into multimodal training datasets for foundation-model research.
    • Built a distributed data platform used by 50+ ML researchers for PB-scale dataset search, curation, filtering and lineage.
    • Designed ingestion, preprocessing and dataset-versioning pipelines for large-scale egocentric and third-person robot data, supporting policy training and learning from human video.
    • Built semi-automated annotation with Segment Anything, CLIP and active learning to speed up manipulation supervision.
    • Defined data schemas, collection protocols and training sets with perception and manipulation researchers.
    KubernetesAirflowFastAPIPostgreSQLGCSSAMCLIP
  3. Real-time perception for autonomous mobile robots operating around forklifts and heavy machinery.

    • Designed a two-stage pipeline, YOLO detection followed by U-Net segmentation with dice loss, for forklift detection under heavy class imbalance.
    • Built distributed multi-GPU training on Ray and PyTorch, and explored ViT transfer learning.
    • Shipped quantized models on NVIDIA hardware for resource-constrained robots.
    • Built 20 TB+ multimodal data pipelines (S3, Docker, FiftyOne, W&B) with automated validation.
    YOLOU-NetRayTensorRTFiftyOne
  4. Perception, sensor calibration and on-robot ML for a fleet of 1,000+ collaborative warehouse robots.

    • Shipped YOLO-based detection of people, robots and forklifts for collision avoidance across the deployed fleet.
    • Invented a point-cloud camera-calibration method (gradient descent on ground-plane estimates) that cut per-site setup from 4 hours to 15 minutes. Patent filed.
    • Built end-to-end ML deployment (GCP, BigQuery, ONNX, TensorRT) serving 1M+ inferences a day, with a 5× GPU speedup.
    • Designed lightweight models (image tokenization, random-forest ensembles) for constrained on-robot compute.
    CalibrationPoint cloudsONNXTensorRTGCP
  5. Industrial research co-advised by WPI faculty: metric scale for monocular SLAM.

    • Recovered monocular SLAM scale from ground-plane detection and camera calibration.
    • Implemented real-time RANSAC ground-plane estimation on 3D point clouds at 90% accuracy.
    • Fused depth estimates with vehicle CAN bus and IMU data for better localization.
    SLAMRANSACSensor fusion

Education

M.S. Robotics Engineering

Worcester Polytechnic Institute · 2017–2019

Thesis-track work on metric-scale visual-inertial 3D mapping with a phone camera and IMU.

 

B.E. Mechanical Engineering, Honors

Rajiv Gandhi Technological University · 2014

Patents

Sensor calibration & edge detection

Patent applications filed at 6 River Systems (2020–2022) for autonomous-vehicle sensor calibration and edge-detection methods for robot safety.