Ankur Goyal

Ankur Goyal founded Braintrust to make evaluation a core workflow of AI engineering: measure each change, identify improvements and regressions, and use those results to guide development. He previously led engineering at SingleStore, founded the unstructured-data company Impira, and led AI work at Figma after its acquisition of Impira. In February 2026, as Braintrust’s founder and CEO, he argued that production AI observability should be treated as core infrastructure.[1][2][4][3]

Braintrust grew from a problem Goyal encountered at both Impira and Figma. A model or prompt change could improve invoice extraction while worsening bank-statement extraction. At Impira, he recalled, the team debated competing approaches for two or three months before building an evaluation prototype with Snowflake and scripts; they shipped the replacement model roughly two weeks later. It still had regressions, but was better in aggregate. Evals gave the team a way to compare trade-offs across cases rather than argue from isolated examples.[2][1]

Elad Gil helped Goyal recognize that building similar internal tooling twice might indicate a wider need. They tested the idea through prospective-user interviews. Those conversations challenged Goyal’s initial preference for an open-source solution: the early users he interviewed wanted something they could install and rely on. Zapier’s Brian followed an interview with an urgent request for help, and Goyal recalled the team using a prototype five days after it was built. Brian’s demanding feedback shaped the developer experience; Coda’s David helped establish demand for a prompt playground.[2][1]

Goyal described Braintrust’s evaluation interface as a declarative specification of data, a task function and scoring functions. The platform handles execution, parallelism and caching instead of requiring each developer to assemble that machinery. Matching TypeScript and Python interfaces let teams move between languages. Using the same function abstraction for evaluation and application execution also lets production logs arrive in a form reusable for eval datasets, reducing the work of converting logs into test data.[1]

Goyal prioritized TypeScript because he saw product builders as the intended users. He recalled building the initial TypeScript SDK over a weekend after Zapier’s Olmo requested it, and credited Olmo’s first use and feedback with shaping its ergonomics. Another early decision was hybrid deployment: Zapier, Coda and Airtable needed their data to stay in their own clouds. Goyal said Braintrust was engineered for that model from the outset, with serverless execution helping contain failures without restarting servers or Linux processes.[1]

Goyal deliberately delayed pricing conversations until early customers began putting Braintrust into production, using their demand as a test of the product’s value. User behavior also changed his view of the playground. Initially uncomfortable with requests to make it an IDE, he accepted that people wanted a durable, collaborative workspace for serious prompt development. In his July 2025 interview, he described the team spending months re-engineering its React state to support that quality bar.[2][1]

Notion’s full-text search through AI logs pushed Goyal back into database development. He said the existing mix of database technologies could not adequately serve the workload as it grew. Goyal and his brother Manu built a benchmark intended to be roughly a hundred times more stressful than Notion’s usage and tested databases and search libraries. In his account, Tantivy, a Rust implementation of a Lucene-style search index, was the only tested option capable of handling the search workload.[2]

Goyal, Manu and Austin, an early Impira engineer, jointly built a purpose-specific database around those requirements. Goyal emphasized its narrow scope: it was designed for Braintrust’s workload, rather than as a general data warehouse replacement. In his February 2026 announcement, he connected the company’s database investment to the growing complexity of AI observability and argued that engineering and product teams needed clarity and accountability as models and applications changed.[2][3]

Before Braintrust, Goyal’s work at MemSQL and Impira gave him experience with database engineering, customer responsibility and the difficulty of finding a repeatable product market.[2][1]

Goyal left undergraduate study to join MemSQL, later renamed SingleStore, and spent roughly five and a half to six years there, eventually running engineering. He credited Adam, a senior database engineer from Microsoft, with teaching him to take broken tests and customer issues seriously. He also recalled a Goldman Sachs managing director warning that a software failure could cost customers their jobs—an encounter that made the responsibility of delivering reliable software tangible.[2][1]

At Impira, Goyal sought to make unstructured data as easy to use as structured data through machine learning. The company eventually focused on extracting information from documents. He later described two difficult lessons: technical solvability did not establish market demand, and selling to nontechnical business users required customer understanding he had taken for granted when serving engineers at MemSQL. Tailoring the product to early customers produced paying users, but too little repeatability to scale the sales process easily.[1][2]

As language models challenged Impira’s existing extraction techniques, the team retooled the product around LayoutLM. Goyal recalled considering the company’s traction, the team’s exhaustion and whether he wanted to raise more capital to continue the work for another five to fifteen years. He decided against it and pursued acquisition, with Gil’s help. During the wind-down, he said the team provided generous refunds, worked closely with customers and wrote code to assist them.[1]

Figma offered a team Goyal admired and a multimodal problem combining textual and visual signals that appealed to Impira’s team. After joining, however, he found its release environment difficult to reconcile with his preference for rapid customer iteration. “I am a creature of shipping,” he said in the October 2024 interview. He described Figma as a poor personal fit while praising its product quality and colleagues.[1]

In 2024, Goyal argued that vector search’s deployment challenges often lay in adjacent application data—permissions, access controls and joins—rather than nearest-neighbor search itself. He similarly distinguished fine-tuning from the outcome customers wanted: better performance through optimization. Fine-tuning, prompt optimization and manually refining prompts were alternative means to that end. He also favored designing systems around improving model reasoning capabilities, while acknowledging uncertainty about how much agent orchestration would move into models.[1]

Goyal’s emphasis on quality extends to how he describes running Braintrust. In 2025, he advocated treating small failures as meaningful signals and described a culture favoring autonomous engineers, designers who code and product managers who prototype, with one company meeting per week. He distinguished those shared expectations from his own habits: although he worked weekends, he said that was not a company-wide expectation.[2]