Summary
Greptile is out to validate every piece of software produced on the planet. As more code is written by AI, the first review pass has to be independent of the tool that wrote it, and it has to be fast. A validation layer that slows teams down defeats its own purpose. Greptile came out of Y Combinator’s Winter 2024 batch and has since raised a Series A led by Benchmark. Its agents build a semantic understanding of the whole repository, so every review reflects the patterns, architecture, and intent behind the code. Together AI provides the inference foundation that lets Greptile run open models at production scale, with the latency its customers expect.
About Greptile
Greptile is an AI code review tool that plugs into GitHub and similar platforms and reviews every pull request automatically. Founded in 2023 by Daksh Gupta, Soohoon Choi, and Vaishant Kameswaran, the company catches bugs, style issues, and logic problems before a human opens the diff.
Greptile’s depth comes from context. It reads the full repository, not just the diff, and draws on intent from tools like Notion and Linear plus patterns learned from how engineers actually work. Because it works across every codegen platform, it sits independently of any one of them. For enterprises where teams code in different tools, that independence matters.
The challenge: independent validation at the speed of generation
Greptile’s review pipeline has to keep pace with how much software gets written. Several requirements shaped its inference needs:
Volume arrives like a waterfall: Greptile takes the first pass over everything a company produces. That scale puts constant pressure on the endpoints behind each review.
Speed is part of the product: Customers care about the end-to-end time of a code review. As coding models get faster, developers expect to ship more, and validation cannot be the thing that holds them up.
Independence shapes the design: Having the agent that wrote the code also review it is a conflict of interest. Greptile orchestrates multiple models so no single one’s blind spots decide what gets caught.
The best model for each task: Different models have different strengths when reading code and catching bugs. Greptile found that open models complement frontier closed models.
The timeline was short: The team was working fast toward shipping the Greptile V5 agent and needed infrastructure that could meet its latency and efficiency requirements without slowing the launch.
“As coding models get better and faster, so too do the expectations of developers in terms of how much they’re looking to ship on a day-to-day basis. As the universal code validator, we want our code reviews to be as fast and performant as possible.” — Rohan, Research Engineer, Greptile
The solution: open models, orchestrated, on fast infrastructure
Many models, one reviewer: The Greptile V5 agent spins up sub-agents to review each pull request, using an orchestration of models to catch the bugs any single model would miss. Open models run alongside frontier models, each doing the work it does best.
Fast from the first demo: Together’s team set up test deployments quickly and confirmed they met Greptile’s latency and efficiency requirements. The process involved almost no friction, which made the decision to move forward easy.
Support that works alongside the team: Together’s technical and solutions engineering teams work hands-on with Greptile day to day.
An inference foundation Greptile can build on: Together runs the open models at production scale, so Greptile’s team can spend its time on review quality and new capabilities.
Results: validation that keeps developers moving
The clearest sign of success for a validation layer is that developers are never blocked waiting on it:
Reviews in half the time: Greptile now ships code reviews that are more performant in half the time.
Latency met from day one: Together’s test deployments met Greptile’s latency requirements from the first demo, so the team moved forward with open models without lowering its quality bar.
A shorter path to launch: Setup involved almost no friction, which mattered with the V5 agent on a tight timeline.
Independent validation across every codegen tool: Enterprises get one tool-agnostic layer, whatever their teams write with.
“Working with Together has enabled us to ship code reviews that are more performant in half the time. For a company like Greptile, inference is the foundation of every product we build.” — Rohan, Research Engineer, Greptile
What’s next: from reading code to testing it
Standard review, human or AI, mostly reads code. Greptile’s T-Rex runs it, finding bugs that reading alone cannot catch. Adversarial path simulation, launching soon, will let T-Rex test how code behaves under realistic failure conditions before it reaches production. It is the next chapter after code review, an extension of it.
As that capability rolls out, Together will keep the speed and reliability in place so Greptile can ship it with confidence. The goal their team describes is a world where every piece of software is validated, and where validation keeps pace with everything being built.
