[Curriculum Vitae]

I am an Applied Scientist II at Amazon Lab126 (Alexa Devices), where I work on perception and scene-understanding systems for on-device and embodied AI. My work brings together synthetic data generation, multimodal foundation models, high-fidelity simulation, and the evaluation methods that tie them together.

Areas of Expertise

  • Synthetic Data & Simulation: diffusion-based generation (ControlNet, custom LoRAs) with VLM auto-annotation, and procedural environment generation in Unity and Isaac Sim — ~30k+ samples driving 5–25% mAP gains on downstream perception
  • Multimodal Foundation Models: post-training vision-language models (Qwen-VL, LLaVA) with supervised fine-tuning, LoRA, and reinforcement learning (GRPO) for on-screen intelligence
  • 3D Reconstruction: monocular, depth-regularized 3D Gaussian Splatting (gsplat) deployed into Isaac Sim with simulated sensors (camera, depth, IMU)
  • Evaluation & Interpretability: DINOv2 / SigLIP embeddings and per-layer CKA / patch-level similarity analysis to quantify the sim-to-real gap and diagnose failure modes
  • Audio–Visual Understanding: multimodal classifiers that fuse audio and image encoder features; audio labeling and balancing with CLAP embeddings

Earlier at Lab126, as an Applied Scientist (2022–2024) I built a Bayesian semantic-grid pipeline reconstructing real environments from Astro RGBD data (>90% accuracy across 30+ environments) and a natural-language-driven environment-generation system (LLMs on AWS Bedrock with CLIP/SBERT retrieval and MILP placement). As a Software Engineer (2020–2022) I built physics-based robot simulation and simulated perception pipelines with sim-to-device parity for Astro.

Before Amazon, I completed my M.S. at the University of Pennsylvania. At the GRASP Lab (with Prof. Kostas Daniilidis) I contributed to RoboNet, a cross-platform real-robot data-collection framework (CoRL 2019), and to Active Semantic Goal Navigation (ICLR 2022); at the PRECISE Center (with Prof. Rahul Mangharam) I built real-time perception and model-predictive control for an F1/10 autonomous vehicle (NeurIPS 2019 demo).

My work has appeared at AAAI, CoRL, ICLR, NeurIPS, and ROMAN, with earlier work featured in MIT Technology Review.