Experience
Where I have worked and what I built there.
Seneca
Software Engineer — Perception
I build the perception stack for autonomous drones — the pipeline that turns raw sensor data into something a flight system can act on. Most of my work sits at the seam between data and deployment: getting training data labeled without a human in the loop, and getting the resulting models running fast enough on hardware that flies.
- Built an automated data annotation and training pipeline, adapting SAM3 and FoundationStereo for labeling, and used it to train YOLO and custom depth models tuned for drone-specific scenarios.
- Designed and deployed an end-to-end fine-tuning pipeline for LiDAR-based human detection running on NVIDIA AGX Orin.
Meta — Reality Labs
Machine Learning Engineer, Computer Vision
My main focus was 3D reconstruction and scene understanding for AR/VR — turning low-quality headset captures into usable 3D, and grounding language in that geometry. Alongside that I worked on stereo depth and segmentation, where the recurring problem was scaling annotation to datasets far too large to label by hand.
- Developed a synthetic data generation pipeline from low-quality multi-view captures using Gaussian Splatting (2D/3DGS), and evaluated 4DGS variants for fidelity and temporal consistency on a CVPR-scale dataset.
- Fine-tuned a language-grounded 3D-VLM for scene understanding and geometry-aware AR/VR panel placement, delivering results on 360K samples with spatial extent and surface normal cues.
- Built end-to-end stereo depth estimation pipelines — pre-processing, inference, post-processing — on a customized NVIDIA Foundation Stereo model, working directly with stereo camera geometry.
- Optimized a hand–object segmentation pipeline, improving GPU utilization by 75%, and integrated SAM3 with box + text prompting into an auto-labeling and QA workflow to scale segmentation on custom datasets.
Vimaan Robotics
Geometric Computer Vision
Worked on the vision system behind warehouse inventory automation — detecting pallets and ground planes from drone and fixed-camera imagery, and calibrating the cameras that everything else depends on.
- Deployed a transformer-based (DETR) detection and segmentation system on cloud for pallet/ground recognition, improving mAP50-95 by 5% by customizing decoder outputs.
- Developed an end-to-end camera calibration module with noise modeling, achieving pose estimation accuracy within 0.25° and 2 cm. Coordinated deployment of TIM551 2D-LiDAR.