Ankur
Goyal

There is a particular kind of frustration that builds slowly, over many sprints, across hundreds of model outputs, until it becomes impossible to ignore. For Ankur Goyal, that frustration arrived while he was running the AI/ML platform at Figma - the company that had acquired his previous startup, Impira, in 2022.

Every team building with AI ran into the same wall: how do you actually know if the model is getting better? You tweak a prompt, you run it, you look at the outputs, you ask someone on the team if they "feel" like it improved. That feeling is not a system. Goyal, who had spent years ensuring that SingleStore's distributed database performed reliably at scale, found the informality of AI quality measurement genuinely alarming.

He founded Braintrust in 2023 because the problem wasn't abstract - he had personally lived it. The company he built addresses the gap between a working AI demo and a production-grade AI product. That gap, it turns out, is mostly evaluation infrastructure.

What Braintrust built is not merely a logging tool. It is an end-to-end evaluation and observability platform: inspect traces, drill into tool calls, track latency and cost in real time, run experiments against real datasets, compare prompts side by side, detect regressions automatically in CI/CD. The platform scores AI outputs using LLMs, custom code, or human feedback - all in one place.

The framing Goyal uses is pointed: "Evals are the new PRD." The vague product requirements document that product managers have always written is being replaced by something measurable, testable, falsifiable. If your AI feature can't be evaluated, it can't be shipped responsibly.

Goyal's instinct for infrastructure comes from a very specific career path. Around 2011, while studying computer science at Carnegie Mellon, he made a decision that required a certain kind of confidence: he dropped out to join MemSQL (later SingleStore) as one of its first employees. This was a company building a distributed, in-memory SQL database at a time when the category barely existed.

He spent roughly eight years there, eventually becoming VP of Engineering. The work was systems-level: horizontal scale-out, distributed transaction handling, in-memory storage engines. He presented the MemSQL architecture at the Carnegie Mellon Database Group - the kind of technical credibility that you earn by building things that actually work under pressure, at scale, for paying customers.

Then came Impira. Founded as an AI-powered platform for document extraction and data management, Impira was the kind of focused, technically-ambitious product that appeals to the engineering-heavy SF startup scene. In December 2022, Figma acquired it - the kind of exit that validates both the founder and the thesis.

Post-acquisition, Goyal led the ML Platform at Figma. It is a peculiar kind of vantage point: deep inside one of the most product-focused design companies in the world, trying to build AI features that meet the quality bar Figma users expect. The evaluation problem kept surfacing, kept resisting easy solutions, kept demanding a purpose-built answer.

He walked back out to build it himself.

What Goyal talks about on X and in interviews is worth following closely, because his frameworks tend to land with practitioners. Two have gotten particular traction.

The first is "context engineering" - the practice of bringing the right information in the right format to language models. It's not prompt engineering, which is about the instructions themselves; it's about the data that surrounds those instructions. Getting this wrong is how production AI products fail quietly: the model has good instructions and bad context, and nobody has the observability to spot the difference.

The second is the "Four Stages of AI Development," a pattern Goyal observed from watching Braintrust's customer base. Organizations don't adopt AI uniformly - they pass through identifiable phases, each with different tooling needs, different failure modes, different readiness for the kind of rigorous evaluation Braintrust enables. The framework is useful precisely because it tells you where your organization actually is, not where it thinks it is.

Braintrust closed a $36M Series A in October 2024, led by Andreessen Horowitz with Martin Casado. The angel list reads like a who's-who of the infrastructure-adjacent AI world: Greg Brockman from OpenAI, Arthur Mensch from Mistral, Guillermo Rauch from Vercel, Simon Last from Notion, Bryan Helmig from Zapier, plus Elad Gil, Greylock, Datadog, and Databricks Ventures.

These aren't passive investors. They are the operators building production AI systems who recognize the evaluation problem from their own experience. When the people who run the most demanding AI infrastructure in the world back your tooling company, it's a signal about what the professional engineering community believes is missing.

In February 2026, ICONIQ Capital led an $80M Series B at an $800M valuation. Matt Jacobson joined the board. Total funding: $121.1M. The positioning: the observability layer for all AI infrastructure.

A detail that doesn't show up in funding press releases: Goyal has contributed directly to DuckDB (specifically the core planner and cast functions) and to Hugging Face Transformers. For a CEO, maintaining hands-on open source contributions is either a statement about how he spends his time or a statement about what he values - probably both.

His philosophy around customers is worth noting: he talks explicitly about "high-bar users" - the engineers who are most demanding, most sophisticated, most likely to stress-test a product and find its edges. Braintrust's customer list (Stripe, Cloudflare, Vercel, Notion) reflects that preference. These are companies that care about correctness and will tell you, clearly and early, when something is not good enough.

Zapier was Braintrust's first paying enterprise customer - a signal that resonates. Zapier's entire business is about reliable automation. Their willingness to pay early, before the product was polished, said something about the depth of the problem Braintrust was solving.