Recent Work

Trained a Franka robot policy to insert a peg into a socket while minimizing contact force, testing whether giving a policy force-sensing actually improves safety and robustness over a geometry-only baseline.

Forked from Isaac Lab's Isaac-Forge-PegInsert-Direct-v0 environment and trained four PPO policy variants (3 seeds each) via RL-Games. The result: force observation is what let the policy learn the task at all. Giving it force feedback roughly doubled success rate (7.3% → 14.2%), while penalizing force without sensing it actually hurt performance.

By Aditya Sugandhi, my co-author on the from-scratch Transformer implementation. An end-to-end workflow for training an object-detection model entirely on synthetic data: generating labeled images with NVIDIA Omniverse Replicator, converting them to KITTI format, and fine-tuning a Faster R-CNN (ResNet-18 + FPN) on NVIDIA DGX Spark hardware.

Keeps the best checkpoint by validation mAP separate from the last, since detection models overfit synthetic data readily, a small but important guard against the sim-to-real generalization gap.

Earned the NVIDIA-Certified Professional: OpenUSD Development (NCP-OUSD) credential, validating the ability to build, maintain, and optimize 3D content-creation pipelines using OpenUSD.

Direct groundwork for InstruX: the scene and data pipelines behind simulation workflows and robot-policy evaluation run on OpenUSD.

A remote-controlled development workspace for running NVIDIA Isaac Sim on cloud GPU instances while driving it from a local Claude Code agent, treating the cloud instance as fully disposable.

Scene work is version-controlled in Git as the source of truth (USD .usda files are text-based and diffable, with assets referenced rather than embedded), while slash commands manage the full instance lifecycle (spin up, save, resume, screenshot) over an SSH tunnel and WebRTC viewer.

A decision and orchestration layer for the robot-policy lifecycle: diagnosing policy failures, generating targeted simulation scenarios, checking deployment readiness, and triggering retraining when deployed systems drift.

Built above Isaac Sim, Isaac Lab, Cosmos, Omniverse, OpenUSD, and ROS2, applying the same production-AI systems discipline from my work at U.S. News to Physical AI.

Built a production agentic RAG assistant for the U.S. News Education B2B platform. The system connects proprietary education datasets, Academic Insights APIs, OpenSearch vector retrieval, LangGraph orchestration, streaming responses, and NVIDIA NeMo Guardrails into a customer-facing analytics workflow.

My work focused on backend architecture, tool-calling reliability, retrieval quality, production debugging, and safe response generation, including guardrails for topic control, PII detection, RAG grounding, and jailbreak prevention, all at low latency.

Built the assistant that powers this site, so recruiters and engineers can ask about my work, projects, skills, and career direction in natural language.

It runs on a serverless backend that streams responses token-by-token, grounded in my own career and project data with a constrained response layer, so visitors can explore my background conversationally instead of reading every page.

Implemented a text-completion model from scratch in Python using the Transformer architecture, inspired by Attention Is All You Need.

Writing the whole thing myself, from tokenization and attention through training, inference, and GPU-aware execution, is what made the mechanics behind modern LLMs like GPT, BERT, and BART click for me.

An autonomous urban-intelligence prototype that turns natural-language questions into live geospatial analysis.

Google ADK agents and the Gemini Live API reason over city data and drive a CesiumJS 3D globe in real time, backed by a Python FastAPI service and a Next.js / TypeScript front end. The focus is agent orchestration and streaming, multimodal interaction over an interactive map. It is a step toward the closed-loop, real-world systems behind my Physical AI work.