- Stera is a full-stack platform for capturing high-fidelity multimodal robotics data on commodity hardware.
- Vidur runs in production, labeling fine-grained actions in raw robot and human video.
- I trained vertical-specific models for clients.
- Indic Chat became a home for community Indic language models and was shown at Google I/O 2024.
- Gajendra is a 7B model that moves between Hindi, Hinglish, and English.
- Dataformer is open-source infrastructure for generating synthetic training data.
- The Hindi datasets run to billions of synthetic tokens, all released in open source.
- Jury Simulator runs virtual US court trials, so attorneys can rehearse a case before they argue it.
Solo projectChat Browser
Jun 2023 - Sep 2023- A browser extension for talking to any website, PDF, or YouTube video while browsing. I built it alone, and a San Francisco startup acquired it.
- A self-checkout vision system went into production at Jackson Hole Airport and Intel stores.
- I also cut object-detection training time by 3-4x within a month of joining.
- 6D pose estimation and segmentation let a robot pick and clean dishes on its own.
- The model trained entirely on synthetic data built in Blender, which saved months of data collection.
- ONNX, OpenVINO, and quantization got it running on edge hardware.