The useful clue arrived disguised as somebody else’s bad frontend. Daksh Gupta and his college collaborators had been building a different startup at Georgia Tech. One friend wrote the interface at startup speed, then chose not to continue after graduation. Gupta, Soohoon Choi, and Vaishant Kameswaran inherited the code. It was hard to navigate, harder to explain, and familiar to anyone who has opened a repository after its original author left.
Their frustration produced a narrow question with a large future hiding inside it: could an AI learn the whole codebase, then answer questions about what the software was doing? The question became a weekend prototype. The prototype became a move. By senior year, the three friends knew they did not want to return to the companies where they had interned. They booked one-way flights to San Francisco.
The company they eventually called Greptile now reviews pull requests with context from across a repository. It checks changes for bugs, applies a team’s engineering standards, and learns from the comments human reviewers leave behind. Gupta’s bet is that the generative coding boom creates its own counterweight. More code does not remove the need for judgment. It increases the surface area that judgment must cover.
The company inside the pivot
Greptile did not begin with code review. At Georgia Tech, the founders entered CREATE-X Startup Launch with an AI shopping assistant called Tabnam. Customer discovery pulled them away from shopping and toward the problem they had personally felt. In July 2023, they shifted their energy to AI tools for software engineers. The early company was called Onboard AI and pitched a broad idea: an AI expert that could understand large codebases.
The name arrived during a late night at the office. The founders typed puns on grep, the command-line search tool, into Namecheap. When greptile.com appeared as available, they jumped up and down. It is an appropriately nerdy bit of founder folklore: a serious infrastructure company named through domain-search wordplay.
The more important naming decision was the product’s job. Codebase chat was impressive, but code review sat inside a recurring workflow with a clear buyer and an obvious cost of failure. A pull request is a little border crossing. New code arrives. Somebody must inspect its papers. Greptile could work there without asking developers to invent a new habit.
A reviewer that remembers the neighborhood
Traditional automated checks are good at fixed rules. A linter can enforce syntax. A test can verify a known behavior. The messy bugs live between files, assumptions, and old decisions. Greptile’s argument is that a useful AI reviewer must understand the neighborhood around a change, not merely the changed lines.
That context can include the codebase graph, organization-specific instructions, related work in Jira or Notion, and patterns learned from prior human comments. If an engineer repeatedly says that authorization must happen before a particular error is returned, the system can learn the rule for later reviews. Over time, the reviewer should sound less generic and more like the team whose code it checks.
In June 2024, with nearly 400 software teams using its API, chatbot, and integrations, Greptile announced a $4.1 million seed round led by Initialized Capital. In September 2025, it announced a $25 million Series A led by Benchmark, alongside a rewritten agent architecture called Greptile v3. By early 2026, Georgia Tech reported that the company served more than 2,000 customers and had raised about $30 million in total.
The customer count matters less than the shape of the thesis. Greptile does not need to pick which coding agent wins. It can inspect code from Cursor, Claude Code, Devin, a human engineer, or some combination of all four. In a market obsessed with owning generation, the company is trying to own the checkpoint.
The founder as public test surface
Gupta’s public persona runs at the same speed as his product philosophy. He describes Greptile as a combination of high standards and urgency, “almost to the point of impatience.” He prefers an in-person office because questions travel across a room faster than they travel through calendars and chat threads. His comments about long workdays and limited work-life balance at an early-stage startup have generated intense criticism, which he has met with more explanation than retreat.
The useful distinction is transparency. Gupta says candidates hear the expectations in their first interview. His position is not that every company should operate this way. It is that an early team trying to reach what he calls escape velocity may choose intensity, and applicants should know the bargain before joining. The argument remains divisive. It also makes his recruiting filter unusually legible.
“Every six months, the job changes.” Gupta’s compact description of company building explains his own path from student engineer to product lead, salesperson, recruiter, fundraiser, and public spokesman. Mastery is temporary; the next stage arrives on schedule.
His candidness shows up clearly in old Hacker News threads. During Greptile’s March 2024 launch, developers questioned permissions, documentation claims, repository limits, interface choices, and reliability. Gupta replied in public. He admitted when the team had not considered something. He acknowledged a “fiddly” interface. When traffic overwhelmed the system, he described an office scramble and kept responding. The exchange reads like an unvarnished customer-support transcript from a company still pouring concrete while people walk across it.
There is playfulness, too. His Y Combinator biography says he interned at Amazon “where I was so good that they named an Amazon locker after me.” Greptile made a branded energy drink. Gupta’s personal site, a nearly empty manifesto about imperfect software, links unexpectedly to a Spotify single. The founder who argues for severe operational focus also appreciates a joke that earns its keep.
Marketing something developers distrust
Developer tools have a peculiar marketing problem. Their users are trained to detect hand-waving. A campaign can win attention and lose trust in the same sentence. Gupta’s answer is to make things that are distinctive enough to remember, then let the product carry the technical burden.
He once doubted that content could bring customers. He later started treating it as the top of the funnel: the large set of people who have heard of the company before the smaller set who might buy. A documentary about a refactor may reach fewer people than an ad, but more of those viewers can remember it. The energy drink works by the same logic. Nobody confuses the can with the API. They remember who made both.
This is founder-led marketing with a useful constraint. The stunt must still rhyme with the company. Greptile’s lizard, its terminal-name pun, its code stories, and Gupta’s direct posts all signal a team steeped in developer culture. Its memorability grows from the same odd little obsessions as the product.
From confidence to compounding
Gupta credits Georgia Tech with two prerequisites: meeting Choi and gaining the confidence to treat a startup as a real career. He credits Y Combinator with a similar trio of effects at a larger scale: ambitious peers, credibility, and accomplished people believing in founders before the founders fully believe in themselves. That confidence, he says, compounds.
The observation helps explain the one-way ticket. The founders did not arrive in San Francisco with Greptile fully formed. They arrived with each other, a personal problem, and permission to keep changing the answer. Their first shopping idea disappeared. The broad codebase assistant narrowed. The narrow reviewer expanded into a validation layer. The throughline is less a product roadmap than a willingness to replace one.
The inspector at the software factory
Gupta’s personal website opens with a severe sentence: “The world is fragile as long as it runs on imperfect software.” The claim is broad, but Greptile’s daily work is specific. Read a pull request. Understand what surrounds it. Catch the mistake before production. Remember the team’s preferences the next time.
The timing is the interesting part. Coding agents promise to produce more software with fewer keystrokes. If they succeed, engineering teams inherit a review problem measured in abundance. The faster the factory runs, the more valuable an inspector becomes - provided the inspector can understand the whole assembly line and not merely the newest bolt.
Gupta is building for that asymmetry. He does not need AI coding to become fully autonomous. He needs it to become common enough that human attention turns scarce. Greptile can then occupy the quiet, consequential moment between “the code exists” and “the code ships.” It is the same moment that began the story: somebody else wrote the frontend, the original context disappeared, and three friends had to figure out what it meant.