Varun

Varun Mohan co-founded Exafunction and led Codeium as it expanded from AI code completion into Windsurf, an editor built for agentic coding. His consequential decisions included using the team’s inference infrastructure to make completion free, building an editor when extensions constrained the experience, and combining proprietary systems with third-party planning models. In July 2025, Google DeepMind announced it was hiring Mohan, co-founder Douglas Chen and some Windsurf researchers.[1][2][4]

Codeium initially offered extensions across editors so developers could use AI without leaving established tools. In the December 2024 interview, Mohan explained that language servers and debuggers mattered to customers who worked in environments such as JetBrains. Windsurf supplied greater control over the experience as the team moved toward agents that could retrieve code, propose changes and execute it.[2]

One limitation was unusually concrete: the team could implement Supercomplete’s refactoring suggestions, but VS Code did not expose the API needed to display them as intended. Mohan recalled dynamically generating PNGs to show the suggestions. Anshul Ramachandran explained the complementary requirement: observing a developer’s trajectory—directories opened, edits made and terminal commands tried—could help the AI infer intent without requiring a complete written specification.[2]

By late 2024, the architecture combined proprietary keystroke-level models, custom retrieval and code-application systems, and third-party models for high-level planning. Mohan argued that embeddings alone could miss the detail required by complex queries, giving “find me all quadratic time algorithms in this code base” as an example. The team ran custom models across large codebases to address that problem; his argument concerned the demands of retrieval, rather than a published comparative score.[2]

Mohan described evaluating Cascade, Windsurf’s agentic component, by removing historical code changes associated with tests and asking the system to complete the resulting incomplete code. Withholding the commit message tested whether it could infer intent as well as retrieve relevant snippets, plan and iteratively reach passing tests. Ramachandran described a separate retrieval evaluation using files edited together in historical commits to form expected sets: software work often requires many distributed snippets, rather than one isolated answer.[2]

Mohan’s late-2024 goal was more complex work with limited human interaction, while preserving the ability to return to the developer when necessary. He advocated snapshots and checkpoints so an agent’s execution trajectory could move forward and backward, potentially exploring multiple alternatives. He identified side effects and latency as obstacles to parallel exploration; these were proposed directions, without a committed delivery date.[2]

On July 11, 2025, Google confirmed to TechCrunch that it was hiring Mohan, Chen and other Windsurf talent to advance its work in agentic coding. Mohan and Chen jointly said they were joining Google DeepMind with some of the team. The arrangement included a nonexclusive license to certain Windsurf technology; Google took no stake or control in Windsurf.[4]

Before Windsurf, Mohan’s work centered on the infrastructure needed to run deep-learning workloads. That earlier work provided serving technology the team later used to build Codeium.[1]

In the March 2023 interview, swyx introduced Mohan as having studied computer science at MIT and become a tech lead manager for autonomy at Nuro. Mohan described managing large-scale offline deep-learning infrastructure there before starting Exafunction. He later recalled internships at LinkedIn, Quora and Databricks, and said he chose autonomous-vehicle work because B2B ETL software did not particularly interest him.[1][2]

Exafunction addressed the difficulty of keeping GPUs well utilized across deep-learning workloads. Its systems virtualized GPU access and decoupled hardware from workloads. The team also built compiler technology, including operator fusion: combining operations so that the fused computation took less time than running the operations separately.[1]

Mohan recalled that the team’s own experience with GitHub Copilot helped establish the practical value of coding assistants. They wrote substantial CUDA and C++, and were surprised when Copilot generated difficult template code in a roughly 500-line file that subsequently compiled. To build Codeium, the team used permissively licensed public code and bootstrapped with open-source models, then shipped its own model with requirements including fill-in-the-middle capabilities, drawing on serving infrastructure already in place.[1]

Free completion followed an engineering and commercial judgment. Mohan said optimization made it feasible, while the team expected to add products and develop paid or enterprise offerings. Giving away every piece of software had not been the original goal. He also explained that model size was constrained by deployment: a larger model might improve quality, but worsen latency or fail to fit the available hardware. Training-compute optimality therefore did not determine the best model to serve.[1]

In 2023, Mohan framed LLM products around latency, quality and correctability—how easily a user could recover useful parts of an imperfect generation. A change spanning ten files could become burdensome if only some edits were consistent. The same principle shaped completion scope: generating an unsolicited else branch while a developer was still writing an if block could add mental load. He favored suggestions that helped developers stay in flow.[1]

The team paired evaluations across languages and codebases with staged A/B rollouts and completion-acceptance signals. Mohan warned about benchmark contamination when evaluation examples entered training data. By 2024, he described using synthetic data and user preference signals—including edits after acceptance—to improve Supercomplete. He supported benchmarks but cautioned against pursuing their final increments at the expense of useful performance across broader tasks: “you probably should find a new hill to climb rather than sort of p-hacking.”[1][2]

Mohan described discussing build-versus-buy decisions with Chen in 2024. He favored buying undifferentiated tools, while preserving competencies such as model inference when they could distinguish the company over time. Outsourcing such work could make the expertise difficult to recover. His qualification was explicit: if the team truly could not keep up, it should buy. Ramachandran also cautioned that some advantages described retrospectively had not been conscious initial decisions.[2]

Mohan and Ramachandran learned enterprise sales through direct customer work before hiring a sales leader. Mohan recalled selling to approximately 30–40 customers by that point; Ramachandran described personally conducting hundreds of deal cycles and continuing to involve engineers in sales and deployments. Mohan valued technical curiosity in a growing sales organization because customers wanted to understand rapidly changing AI systems. He credited the engineering team’s versatility for the company’s transitions and valued colleagues with high standards who could expose weak ideas. His stated launch policy in 2024 was equally direct: “we will never launch with a waitlist.”[2]