Participated · Equity · Aug 2026

Perseptis began as Whiteboard.ai and pivoted to a structured B2B product that automates creation of high-quality business slides and commercial proposals. The founders — former Apple and McKinsey engineers — developed the product to cut the time consultants and small teams spend preparing decks. The company offers a freemium model with paid tiers for advanced features and reports usage by more than 200 companies, including consulting firms, startups, VC funds, investment banks, PR agencies and MBA students. Perseptis emphasizes slide quality meeting MBB-style standards and says users can reduce hours of slide work to minutes. Current plans include launching an API, building integrations and completing security reviews to support corporate customers as the company scales. The team is focused on product-market fit, sales-driven growth and maintaining quality while expanding adoption.

Participated · Seed · Jul 2026

SkyPilot builds a managed AI compute platform (SkyPilot Platform) atop its open source project to help organizations operate large-scale AI workloads across hyperscalers, neoclouds, Kubernetes clusters, and multiple accelerator types. The platform supports development, agentic workloads, training, reinforcement learning, inference, evaluations, and multi-cluster production serving while offering fleet-wide GPU health monitoring, automated remediation, team and quota management, single sign-on, and SOC 2 compliance. The SkyPilot open source project has surpassed 14 million downloads and attracted over 280 contributors, and the company says hundreds of organizations use the project. Private preview customers have used the platform to manage over 10,000 GPUs and support more than 200 researchers per organization, with some reporting up to 20x performance improvements versus the open source tooling. SkyPilot was founded by Berkeley researchers Zongheng Yang, Zhanghao Wu, and Romil Bhardwaj alongside Ion Stoica and Scott Shenker, and the company launched from stealth alongside its $20M seed raise to fund product, engineering, and go-to-market expansion.

Participated · Series A · Jul 2026

Bespoke Labs builds realistic environments that simulate company infrastructure—codebases, microservices, logs, tickets, email, and Slack—to train long-horizon AI agents. The company emphasizes a research-first approach, staffed by research scientists and engineers rather than contractor-driven teams. Bespoke Labs contributes to open benchmarks and datasets such as Terminal-Bench and OpenThoughts and develops techniques like the Genetic‑Pareto Agent Optimizer (GEPA) for automated prompt and policy search. Its customers use the platform to train, evaluate, and optimize agents, and the company focuses on environment curation, synthetic environment generation, and snapshotting real-world infrastructure. Founded in 2024 by Mahesh Sathiamoorthy and Alex Dimakis, Bespoke Labs is based in Mountain View and recently raised $40 million to scale research, infrastructure, and business efforts.

Participated · Series B · Mar 2026

depthfirst develops an AI-native security platform called General Security Intelligence that reasons across code, infrastructure, and business logic to identify complex vulnerabilities and provide ready-to-merge fixes. The company introduced dfs-mini1, a security-focused model initially trained to secure cryptocurrency smart contracts and post-trained via reinforcement learning in security-specific environments. depthfirst reported that dfs-mini1 outperformed frontier models on OpenAI EVMBench while operating at 10x to 30x lower cost and that its approach appears to generalize to other security tasks. Since making products generally available in late 2025, depthfirst has partnered with customers including ClickUp, Lovable, Supabase, incident.io, and Moveworks. The platform has demonstrated strong developer adoption metrics, with 80% of its fix recommendations accepted and merged. depthfirst plans to use new funding to train more specialized models, grow its AI research team, and accelerate enterprise adoption.

Participated · Equity · Jan 2026

Scholé AI is an AI-native learning engine spun out of research at EPFL and UC Berkeley that tailors short, interactive lessons to each employee’s role, tools, and day-to-day tasks. The platform continually adapts difficulty, format, and content in real time while grounding instruction in a company’s own materials, enabling practical, context-aware upskilling. An early focus is teaching employees how to apply AI in their daily work, with Harvard-co-developed AI-intensive courses already used by learners at hundreds of organizations including Bank of America, NASA, Oracle, Microsoft, and Apple. Forbes named the program the best way to learn AI agents for 2026. Scholé AI is currently piloting with Swisscom, Decathlon, and Coop, and operates from Lausanne and San Francisco. The company was founded by Dr. Vinitra Swamy (CEO) and Dr. Paola Mejia (CTO), who each hold over a decade of AI-in-education research experience. To date, Scholé AI has raised $3 million to expand enterprise deployments, grow its team, and further develop its adaptive learning platform.