Participated · Seed · May 2026

Tensormesh builds and sells Tensormesh Inference, a SaaS inference platform that uses KV caching to store and reuse computed model state to avoid redundant GPU computation. The platform delivers up to 10x reductions in latency and GPU spend by serving repeated context from cache and reports cache hit rates, token-level cost breakdowns, and other performance metrics in real time. Tensormesh offers serverless, OpenAI-compatible API access for immediate use and reserved deployments for enterprises needing dedicated capacity and SLAs. The company also maintains and contributes to the open-source LMCache project (over 8,000 GitHub stars) and highlights integrations across vLLM, SGLang, TensorRT, and multiple cloud vendor offerings. Tensormesh is led by CEO Junchen Jiang and was founded by faculty, PhD researchers, and alumni from the University of Chicago, UC Berkeley, and Carnegie Mellon. The company has raised $24.5 million in total funding and markets a pricing model that bills cached input tokens at $0 for serverless deployments, with well-optimized customers regularly achieving cache hit rates above 70%.

Participated · Series A · Jan 2026

Spun out of a UC Berkeley research initiative in 2023, LMArena runs public leaderboards where 5 million monthly users vote on head-to-head model responses drawn from providers such as OpenAI, Google, Anthropic and xAI. The site has logged roughly 60 million conversations a month across 150 countries, generating a unique, crowdsourced benchmark that model developers closely watch. In September it launched AI Evaluations, a paid service through which enterprises, model labs and developers commission bespoke testing, giving the company an annualized consumption rate (ARR) of about $30 million just four months after launch. LMArena positions this data network as the foundation for future products aimed at systematic, large-scale model assessment. The company contends its neutrality despite criticism that partnerships with major model makers could bias rankings. Founders Anastasios Angelopoulos and Wei-Lin Chiang previously funded the project with grants and donations before commercializing. To date the startup has raised $250 million in equity financing and was recently valued at $1.7 billion post-money.

Led · Seed · Nov 2025

Voio builds frontier vision-language models that interpret complete radiology exams and generate high-quality draft reports, unifying image viewing, reporting software, and electronic health record data in a single workflow. Its first model, Pillar-0, is open source and achieved a 0.87 AUC across 350+ findings, outperforming comparable systems from Google, Microsoft, and Alibaba by 10%–17%. The founding team’s earlier AI tools are already validated in 92+ hospitals across 30 countries and have analyzed more than two million mammograms. By reducing context-switching and automating routine reporting tasks, Voio aims to alleviate radiologist burnout, shorten turnaround times, and push imaging from reactive diagnosis to proactive, predictive care. The company plans to scale its technology to multimodal agentic workflows and to support independent benchmarks that bring transparency to radiology AI. Co-founded by UC Berkeley and UCSF faculty Adam Yala, Dr. Maggie Chung, and Trevor Darrell, Voio combines academic research rigor with commercial product development. Its open-source strategy encourages external validation and community contribution, positioning the platform as an infrastructure layer for next-generation radiology.

Led · Seed · Mar 2025

Arcade.dev provides a secure action layer that enforces authorization, executes agent actions, and provides governance and auditing for production AI agents. The company authored the MCP authorization specification and says it runs agent actions in production across major enterprises, including a top U.S. bank, Prosus, and LangChain. Arcade reports tool call volume is up 25x in six months and maintains a catalogue of over 8,000 MCP tools purpose-built for agent usage. Its platform is designed to eliminate standing permissions and reduce failed actions and token costs. The team includes builders from identity, data, and integration layers at Okta, Redis, MongoDB, Snowflake, and Airbyte. With $72 million in total funding after the Series A, Arcade plans to expand product development, grow its ecosystem, and scale hiring to support enterprise deployments.

Participated · Series C · Apr 2024

Founded in 2022, Perplexity provides an artificial-intelligence search and answering platform that pulls content from verified sources to give concise, citation-backed responses. The product is positioned as a next-generation alternative to traditional search engines, allowing users to ask open-ended questions and receive conversational, source-linked answers. To deepen user engagement, the company releases curated experiences such as the newly launched Cristiano Ronaldo hub, which lets fans explore never-before-seen images, interactive pitch visualizations of goals, and detailed career statistics. This strategy highlights Perplexity’s aim to blend authoritative data retrieval with immersive, vertical-specific content. Ronaldo’s endorsement is expected to boost the platform’s global visibility, particularly among sports enthusiasts. While no revenue or user figures were disclosed in the article, the partnership signals Perplexity’s plan to attract high-profile collaborators and embed its AI tools in mainstream culture.