Soham Ganatra

Soham Ganatra cofounded Composio with Karan Vaidya. His AI work centers on the interface between agents and software: managing authentication, representing APIs in forms models can use, and making tool calls more reliable. In his March 2024 account of that work, he proposed managed integrations so developers could concentrate on building agent applications.[3][5]

Ganatra distinguishes agent integrations from interfaces designed for people. In a September 2025 interview, he explained that a person creating a CRM lead might select a category from a dropdown; an agent instead needs an API representation, such as a JSON schema. That changes how developers expose the same underlying operations. It also requires attention to context limits: handing a model an excessively long email can overwhelm the information it can consume.[7]

His April 25, 2024 article, “Improving GPT 4 Function Calling Accuracy”, credits Krishna Pagare and Sawradip Saha as coauthors. The authors built a small ClickUp benchmark with 50 prompt-and-solution pairs covering eight functions, testing gpt-4-turbo-preview at temperature zero. They flattened nested parameters into names such as features__due_dates__enabled, preserving parent information in the names and descriptions, then experimented with system instructions, schema summaries, clearer function names and descriptions, and examples.[4]

The authors reported scores of at least 75% in every run after combined schema and prompt changes. Adding full function-call examples subsequently degraded performance, while adding example parameter values later helped. These were results on their integration benchmark, rather than a general measure of GPT-4 reliability. The article gives inconsistent baseline figures—30% averaged over three repetitions in the methods, but 36% in its conclusion, which reports a final average of 78%.[4]

In July 2025, Ganatra outlined a broader ambition: shared infrastructure through which one agent’s experience with a tool or workflow could benefit other agents. His illustrative example was an agent recruiting a founding engineer and learning useful search and screening heuristics through feedback. He explicitly qualified the vision: “We're nowhere close.” Capturing tacit knowledge, learning from edge cases, and allowing skills to evolve remained problems the team was trying to solve.[6]

Ganatra described MCP in the September 2025 interview as another way to consume Composio’s existing integrations, rather than a change to their underlying quality. Its value, in his view, was extensibility: users could choose connections—such as linking a coding tool to Jira or Notion—beyond the integrations an application’s developer had chosen to provide natively.[7]

Before Composio, Ganatra had worked on chatbots and experienced integration problems inside another startup. Recalling that history in September 2025, he said he started a consumer AI business in 2017, moved toward enterprise chatbot customers, and exited in 2019 in what he called a small exit. He subsequently joined fraud and risk-management company Bureau as a founding engineer, later taking on product and engineering management responsibilities.[7]

Ganatra recalled leaving his job around April–May 2023 after weekend coding experiments gave him confidence to pursue automated software integrations. The initial team tried to build coding agents that could generate the custom transformations needed to connect different systems. After roughly six months, including a proof of concept with Glean, he concluded that the models were not capable enough. The team changed direction: it productized the integrations it had built for its own agents, releasing the resulting product around late March 2024.[7]

After the pivot, Ganatra deliberately sought customers who did not know him. He said friendly contacts could accommodate a product that did not meet their needs and delay honest feedback: “I just want to find people who don't know me and make sure that they like my product at the end of the day.” He recalled averaging more than ten customer calls a day for an extended period, using developers’ accounts of what they were building to shape the roadmap.[7]

Ganatra also used programmatic landing pages to discover demand for particular combinations, such as LangChain with HubSpot. He said visits signaled which integrations people wanted, but some pages advertised connections that were not available. After users objected, the team removed those pages.[7]

His reliance on feedback after launch coexists with skepticism about asking customers to validate unfamiliar AI ideas in advance. In September 2025, he advised founders with strong intuition to build and demonstrate a product because prospective users might not understand its potential from a description. He remained cautious about Composio’s own product-market fit, saying rapid changes in AI made lasting fit difficult to judge: “Never, I still don't know if we do have a true product market fit.”[7]