Varun Shenoy

Varun Shenoy is a cofounder of Long Lake, where he works on bringing AI into services businesses. His central concern is AI diffusion: turning capable models into systems that help employees complete real work in property management, architecture, human resources, and other industries. Long Lake acquires and partners with these businesses, putting the team building the technology inside the organization responsible for its results.

Early work in medical AI and model interfaces

Shenoy studied electrical engineering at Stanford and previously worked at Baseten. His earlier career also included medical imaging work at the U.S. Department of Veterans Affairs and work at Berkeley AI Research and Doss. He coauthored the 2018 research project Deepwound, which combined an ensemble of convolutional neural networks for classifying postoperative wound photographs with a mobile application intended to help patients track recovery at home.

His language-model projects explored practical interfaces for unfamiliar capabilities. GraphGPT, a weekend experiment, turned unstructured text into visual graphs of entities and relationships. Users could supply a movie synopsis or video transcript, inspect the resulting network, and extend it through successive queries. Shenoy also built the initial prototype of HealthGPT, an experimental application based on Stanford’s Spezi framework that let users ask natural-language questions about their Apple Health data. It became an open-source project of the Stanford Biodesign Digital Health team; his credited contribution was the initial prototype.

Faster inference and structured output

At Baseten, Shenoy worked on the mechanics of efficient model inference. His transformer inference guide with Philip Kiely explains how GPU compute capacity, memory bandwidth, and a model’s arithmetic intensity—the amount of computation relative to data movement—shape hardware and batching choices. Their Stable Diffusion XL optimization guide examined fewer sampling steps, changes to classifier-free guidance, compilation, and memory-efficient components. Both guides connect an understanding of performance bottlenecks with empirical benchmarking.

In 2024, Shenoy and Alex Derhacobian developed Super JSON Mode, a framework for accelerating structured output generation. It separates independent fields into smaller requests and generates their values in parallel, avoiding the need to produce an entire JSON object sequentially. An address, for example, need not wait for a bedroom count. Dependencies limit the approach: fields whose values rely on one another cannot safely be generated as independent tasks.

Building agents inside services businesses

Long Lake extended this work on software into the operating conditions that determine whether people use it. In his 2026 account of deploying AI in services businesses, Shenoy explains why the owner-operator model matters: when an agent fails, the unfinished task remains Long Lake’s operational problem. He focuses on work whose essential knowledge often resides in experienced employees’ heads and older software—closing the books when receipts are missing, coordinating vendors to repair a roof, or collaboratively interpreting a construction blueprint. These exceptions are central to the job, and a plausible response alone does not establish that the work is done.

• Earned agent autonomy: Shenoy’s approach to earned agent autonomy starts with an answering copilot, progresses to an interactive agent that uses tools, and then moves toward background execution, long-running work, and a proactive coworker. Each step requires adequate model capability and close iteration with employees. Background agents can also respond to events rather than wait for a user’s request: completing one task might place work in a queue that triggers an agent to offer advice. The desired coworker is the end of this progression, with greater responsibility earned through reliable results and employee adoption.
• Asynchronous knowledge work: He sees asynchronous knowledge work as a particularly important design problem. Engineers already have ways to separate and review parallel work: a coding agent can run in an isolated sandbox, build and test its changes, and return a pull request. Services employees often work serially, handling an inbox one message at a time. Shenoy asks what equivalent structures would let them delegate several tasks and inspect the results later. He also explores representing knowledge work as code to use models’ existing coding strengths. These are open design questions; the interfaces that work for software engineering may need to change substantially for architecture or property management.
• Outcome-based evaluations: His outcome-based evaluations connect agent improvement to what happens in the business. Employees and agents working together produce traces of tool calls, mistakes, corrections, and other obstacles. Long Lake uses these traces to build evaluations tied to completed work: did the roof get repaired, or did the books get closed? Feedback includes explicit ratings and notes, as well as differences between AI-generated material and what employees ultimately submit. Improvement benchmarks become regression tests, allowing the team to check earlier capabilities as it tackles newly encountered tasks. Shenoy also describes internal model post-training on operating data and agent customization for the different practices of companies, individual employees, and clients.

Software-service co-design and employee adoption

Shenoy calls the adoption process software-service co-design. Employees’ use of an agent supplies feedback that can improve it; a more useful agent gives employees a reason to use it again. Starting that loop requires deliberate work. He favors embedding products in familiar tools such as Excel, enterprise resource planning systems, or design software, reducing the effort required to fit AI into an established routine. He also argues strongly for visiting employees in person: observing their work, teaching tools one-on-one, and learning about difficulties through informal conversations as well as training sessions. For him, the people and processes around the software are part of its design, and adoption is part of the same improvement loop as model learning.