Harjot GillCo-founder & CEO

CodeRabbit is a leading provider of AI-powered code reviews. As developers adopt agentic coding tools, the volume of code grows exponentially, shifting the bottleneck from code generation to code validation. CodeRabbit provides automatic code reviews, finds bugs, and explains code changes to developers on every pull request.

I was a teaching assistant at the University of Pennsylvania when I met my future network observability startup co-founder. We spun Netsil out of academic research, focusing on the chaos of the early cloud, Docker, and Kubernetes transition, before ultimately exiting the business to Nutanix.

History rhymes. During the cloud computing boom, deployment became incredibly easy, which shifted the bottleneck downstream to observability and management. Today, generating code is incredibly easy, which shifts the bottleneck downstream to code reviews, validation, and explaining AI outputs to human teams.

During my second startup, GPT-3.5 and Copilot arrived, and it was clear that coding was going to be automated. Managing a remote team across Eastern Europe, India, and the United States, code reviews were already a bottleneck. I knew AI reviews would be the inevitable next big developer workflow.

AI agents can generate massive pull requests instantly, but senior developers are stuck spending all their time reviewing them. Humans cannot read code as fast as AI writes it. Developers need tools that explain the intent, structure, and potential blast radius of changes to build trust.

We designed CodeRabbit to trigger automatically via a web hook the moment a developer opens a pull request. There is no manual human trigger or chat input required. The code review runs silently in the background and delivers high-quality feedback directly inside GitHub or GitLab.

When a web hook fires, CodeRabbit clones the code, creates a live code graph, and runs static analysis tools. We run map-reduce style workflows to gather context across multiple models, prune irrelevant details, and then pass the consolidated context to reasoning models to identify bugs.

We do not let users choose a model. Instead, we use an ensemble of models behind the scenes. We use smaller models for context assembly and semantic analysis, and reserve the most expensive, reasoning-heavy models for the final review. This keeps our token consumption highly efficient.

Open source maintainers perform thankless work, and we wanted to give back. Making CodeRabbit free for public repositories allowed corporate developers to see our product in action on GitHub. It also provided a feedback loop that helped us rapidly train and improve our context assembly.

Unlike single-player IDE tools that only remember local chats, CodeRabbit has a centralized global memory layer. Every developer interaction or instruction helps train the system. The knowledge compounds over time, meaning the system quickly learns the unique quirks and conventions of the enterprise codebase.

We have moved to smaller pods called war rooms that combine product managers, designers, developers, and developer relations. Because AI handles the low-level coding, our PMs and designers can prototype and edit live code directly, reducing hand-offs and boosting shipping velocity.

We look for people with excellent systems design knowledge, deep passion, and good taste. While AI can write code, humans must still formulate the right hypotheses, ask the correct questions, and trace bugs. Developers must move from line-by-line coding to higher-order systems architecture.

Code editors are single-player, local workspaces. As coding agents run for days to solve complex tickets, the workflow shifts to command-line interfaces, desktop applications, and central collaborative surfaces. Just like steering wheels on self-driving cars, traditional editors will eventually disappear.

It requires significant engineering work to optimize our backend code graph, sandbox environments, and model routing. By using semantic analysis to prune the context passed to reasoning models, we ensure that we only consume expensive tokens when absolutely necessary to deliver value.

We are glad to have competition because it validates the market opportunity. CodeRabbit differentiates itself by focusing on the central review choke point. A coding agent cannot grade its own homework; you need an independent, team-based validation layer to build trust and maintain consistency.

Seniors will spend less time on low-hanging syntax errors and more time on high-level architectural design and planning. They will act as team directors, defining the destination and guardrails for AI agents, rather than manually writing and reviewing every line of code themselves.

We want to expand CodeRabbit into a collaborative AI platform for the entire engineering team. Beyond code reviews, we are building integrations for Slack, Jira, and Linear to assist with team planning, issue tracking, and incident response from a central engineering vantage point.