Bhashwar Ghosh

Engineer, San Francisco · bhashwarghosh@gmail.com · LinkedIn · GitHub · Resume

I spent four years at NVIDIA on Metropolis, the video analytics platform. It takes real time streams from hundreds of cameras and sensors and turns them into structured data you can query. I built the analytics pipelines, the vision language and 3D perception systems that produce that data, and a deep video research agent that answers questions over long video. Four US patent applications filed.

Now I am working on training robot foundation models, mainly on inference and data infrastructure.

Now

Aelior Labs / Founder

Robots for physical work in changing environments, instructed in plain language, with one person supervising several remotely.
Working on the Stanford BEHAVIOR Challenge 2026 with a small team. 100 long horizon household manipulation tasks in Omniverse from RGBD and proprioception, against pi 0.5 and GR00T baselines.

NVIDIA, Metropolis

Senior Software Engineer / Jul 2021 to Sep 2025

Built a deep video research agent that answers queries over video, using vision language models and graph RAG for retrieval.
Built the systems that compute vision analytics from real time streams of hundreds of cameras, audio and other sensors, and emit them in a schema built for retrieval.
Worked on 3D vision language models, including SpatialRGPT, answering spatial questions about distance between objects, whether an object fits a given space, and object dimensions.
Combined LLMs, VLMs, 3D VLMs, graph and hybrid RAG, grounding SAM and depth models into agentic analytics systems in LangGraph, and optimised those workflows for latency and performance.
Worked on a warehouse management AI agent that was shown in the NVIDIA GTC 2025 keynote.
Promoted twice, IC1 to IC3, reaching Senior Software Engineer.

AI and agents

AI Video Editor

Upload A-rolls and B-rolls, and it edits the video from a prompt.

Live Demo

Viral Video Generation Engine

Researches what is going viral, reverse engineers the viral element, and generates short form video from it.

Live Demo

Praxity AI

Business insights for CXOs without writing SQL.

Code Demo

RupantarAI

Virtual try-on for apparel e-commerce.

Live Code Demo

AI Version

Discord based chat service where you create your own multimodal AI clone.

Live Code Video MVP clip

Swiggy AI Bot

Order from Swiggy through a bot.

Code Demo

Avatarify

Avatar generation platform.

Live

Web3, earlier work

Decademia

Marketplace for funding, collaborating on and transacting early stage scientific research projects.

Live Code Contract

DDW App

Web3 dating platform built around chatting before matching.

Live Code Contracts Demo

DDW Ads

Advertisers choose keywords, and users are indexed against them from their likes and interests.

ETHGlobal Code Contract

FutureFlows

Buy and sell shares based on your forecasts for the future.

Live Code Contract

ChadCard NFT

Refer to earn NFT, with a chance to earn up to 50 percent on each referral.

Live Code Contract

FemPure

Decentralised e-commerce platform made for women.

Code Contract

Swingby Protocol, freelance

Unstoppable Domains integration with the Swingby DeFi protocol.

Pull request

Patents

Four US patent applications, all assigned to NVIDIA Corporation. Filed, not yet published.

Object behavior identification in video using artificial intelligence models and historical object interaction data. Application 19/544,956, filed Feb 2026, first named inventor.
Performing operations using predicted outputs of a neural network. 2024, first named inventor.
Detecting movement of objects using one or more neural networks. 2024.
Mixture of physical AI agents. 2025.

Education

IIT Madras, B.Tech in Engineering Physics, 2017 to 2021.
Stanford Professional Education, Artificial Intelligence, 2025. Reinforcement Learning (XCS234), Deep Generative Models, Machine Learning with Graphs, Natural Language Processing.