Greptile.

There is a small, universal dishonesty at the heart of software engineering, and it is this: when a colleague opens a pull request and asks you to review 600 lines of changes across nine files, you do not read all 600 lines. You skim. You look at the parts you understand, you type "LGTM" - looks good to me - and you click approve, and the code ships, and sometimes it breaks in a place you never looked. Everyone does this. It is not laziness so much as arithmetic: reading code carefully is slow, and shipping is the job.

Greptile, a San Francisco company founded in 2023, is a bet that this particular dishonesty is a business. Its product is an AI agent that reviews pull requests - but the interesting part, the part that makes it a company and not a browser extension, is what it reads first. Before Greptile comments on a single line of a change, it ingests the entire repository and builds a graph: every file, every function, every dependency, mapped as a structure it can walk. Then, when the change arrives, it does not just look at the diff. It traces the change outward through that graph, follows the function calls into other files, checks the git history, and asks a question most reviewers are too busy to ask: what else does this touch?

This is the whole pitch, and it is a surprisingly narrow one. Most AI code tools read the diff - the highlighted red-and-green block of what changed - and reason about it in isolation. That is fast and it is also how you miss the bug where a renamed field three files away quietly breaks a caller nobody remembered. Greptile's claim is that context, not cleverness, is the scarce resource, and that if you give a model the whole codebase to reason over, it starts catching the class of bug that human reviewers skim right past.

The company was started by three Georgia Tech alumni - Daksh Gupta, Soohoon Choi, and Vaishant Kameshwaran - who went through Y Combinator's Winter 2024 batch and raised a $4 million seed round led by Initialized Capital. The original idea was broader and squishier: help large language models understand large codebases, a sort of semantic search for engineers. The version that stuck is narrower and more useful, which is often how this works. Code review is a task everyone hates, does badly, and cannot skip. If you can automate the part everyone dreads, you get to sell to every engineering team that ships software, which is most of them.

By late 2025, Greptile said it had reviewed more than a billion lines of code across over 2,000 teams, with a customer roster that includes Brex, Substack, Retool, Klaviyo, NVIDIA, Scale, PostHog, and Zapier. In September 2025 the company raised a $25 million Series A led by Benchmark - the venture firm's partner Eric Vishria backed the round - at a reported valuation of around $180 million. Benchmark is famously concentrated and famously early; it is the kind of check that says a firm thinks the category, not just the company, is going to be large.

The false-positive problem

Here is the counterintuitive thing about AI code review, and the thing that separates a demo from a product: the enemy is not missed bugs. The enemy is noise. An AI reviewer that leaves fourteen comments on your pull request, eleven of which are wrong, is worse than no reviewer at all, because engineers learn to ignore it - and an ignored reviewer catches nothing. In independent benchmarks, Greptile posted a high bug-catch rate (around 82%) but also a higher false-positive count than some rivals, which is exactly the tension the whole category is fighting over. Catch more, and you also cry wolf more.

Greptile's answer is to make the AI show its confidence. Every finding comes with a score - a number that says, roughly, "how sure am I about this." The point is not that the machine is always right. The point is that it tells you when it isn't, so you know which three comments to actually read. It also learns: the system watches the feedback engineers leave on its suggestions and folds a team's coding standards, in plain English, into future reviews. Tell it once that your team hates a certain pattern, and it stops flagging things that aren't that and starts flagging things that are.

The workweek that went viral

You cannot write about Greptile without the tweet. In 2024, CEO Daksh Gupta - then 22 - posted that he had started telling job candidates in the first interview that Greptile does not offer work-life balance: an 84-hour week, roughly 9 a.m. to 11 p.m., six days, sometimes seven. The post was viewed around two million times, hit the front page of Reddit, and earned him, by his own account, a spectrum of responses ranging from agreement to death threats. It also earned him applications. Gupta later said the mode is not permanent and not sustainable - that as the company matures it will hire more experienced people with families and naturally soften - which is the sort of thing that is both a genuine belief and a useful thing to say after the internet has yelled at you.

The episode is worth noting less for the culture debate than for what it reveals about the company's temperament: a small team, roughly 21 people, operating with the conviction that in enterprise software there is no prize for second place. That conviction shows up in the product cadence. Greptile v3, shipped in 2025, moved the system from a single-pass reviewer to an agentic one - agents that run in a loop, with tools to search the codebase and recall learned rules, recursively chasing nested function calls until they understand what a change actually does. The company reported it performed dramatically better than the version before it. There is also TREX, a runtime validation agent that writes and runs its own tests against a pull request, which is a genuinely odd and interesting idea: a reviewer that does not just read your code but tries it.

Who buys it, and why

The business model is the familiar per-seat SaaS one - roughly $30 per engineer per month on the Pro plan, with reviews metered beyond an included allotment, and enterprise deals above that. The more interesting commercial move is self-hosting. A large share of the teams who most want AI code review - banks, health systems, defense-adjacent firms - are precisely the teams contractually forbidden from shipping their source code to a third party's servers. Greptile lets those customers run the whole thing inside their own AWS or an air-gapped environment, using their own LLM providers, with SOC 2 compliance on top. It is a less glamorous feature than agentic reasoning, and it is probably worth more revenue, because it unlocks the customers who have the money and the compliance officers.

Greptile competes in a suddenly crowded lane - CodeRabbit, Graphite, Qodo, and the review features creeping into tools like Cursor - all racing to own the moment between "code written" and "code merged." That the category exists at all is a consequence of a larger shift: as AI writes more of the code, the bottleneck moves downstream to reviewing it. If machines are going to generate a rising share of pull requests, somebody - or something - has to check them. Greptile's wager is that the checker should understand the whole codebase, and that the green lizard in the corner of your pull request is going to read the part you were going to skim.