Sonder
ActiveAI-nativeOn-device vision models that watch your screen, learn your workflows, and act proactively—3× faster at half the baseline model size.
What it does
Sonder is an AI lab building on-device vision models that watch people’s screens, learn how they work, and act before they think to ask. Its models turn the context and workflows they observe into shareable skills that teammates’ agents can use. Sonder is built for enterprises that want to own their intelligence and data. The team is post-training small vision-language models and has already achieved 3× faster inference at half the size of the baseline on consumer-grade laptops. They built a desktop RL environment to train these models, keeping it in-house despite outside licensing interest. Sonder is co-founded by Riyan and Ajit who met 2 years ago building and deploying some of healthcare’s earliest agentic systems at Anterior (Sequoia/DeepMind founders backed). CEO Riyan is a 2nd-time founder (FutureClinic — YC/SV Angel) and has been post-training models since the GPT-1 era. CTO Ajit studied CS at University of Hong Kong at age 16. He was previously a senior engineer at Anterior
Its site says now · captured 2026-09-19
Building Your Digital Twin
Next to it in Agent infrastructure · Desktop workflow understanding for agents
- MinicorY Combinator X26
Scalable windows desktop automations
- SourcebotY Combinator F25
Helping humans and AI agents understand massive codebases
- CuaY Combinator X25
Give every agent a cloud desktop
- Cashew LabsY Combinator X25
Recursive self improving harnesses and models
- CloudglueY Combinator S24
The video context layer for AI.
- AgentCata16z speedrun SR007
Session replay and observability for AI agents across MCP, Claude Connectors, and ChatGPT Apps.
- AmuletY Combinator S26
High performance file system for agents
- ChatOverflowa16z speedrun SR007
The network where AI agents learn from each other