This article is based on reporting originally published by VentureBeat in August 2019, covering DeepCode’s $4 million seed round and the technology behind its AI-driven code review platform.
Why Automated Code Review Matters
As software becomes central to virtually every industry, the volume of code that development teams must produce, review, and maintain has grown enormously. More code means more opportunity for bugs, security vulnerabilities, and technical debt to accumulate — and traditional code review processes struggle to keep pace.
Code review sits upstream of software testing. Where testing examines how software behaves from the outside, code review inspects the structure and logic of the source itself, ideally before changes are merged or deployed. In practice, this involves developers, peers, and managers collaborating to catch obvious errors early — a process that scales poorly as codebases grow.
Static analysis tools such as Lint have long been available to help, but they tend to focus on a narrow band of stylistic and formatting issues rather than deeper logic errors or security flaws. That gap is what Swiss startup DeepCode set out to close.
What DeepCode Does
DeepCode applies machine learning to the code review problem. Its models are trained on billions of lines of code drawn from public open source repositories — primarily GitHub — learning patterns from real-world projects, including earlier versions of files and the subsequent fixes that developers applied.
This approach allows DeepCode to go beyond syntax checking. According to cofounder and CEO Boris Paskalev, the system attempts to infer the intent behind a piece of code, not just whether it compiles or conforms to style guidelines. That means it can flag more serious issues such as:
- Cross-site scripting (XSS) vulnerabilities
- SQL injection risks
- Logic errors and known bug patterns
- Security anti-patterns learned from open source fixes
The training pipeline works by examining how bugs appeared in historical versions of code and how they were subsequently resolved. This creates what DeepCode describes as a live knowledge base of known bugs and their solutions, which the model applies when reviewing new code submissions.
Beyond pattern matching on known issues, DeepCode also uses predictive and inferential algorithms to surface problems in code that may not have been explicitly fixed in any training sample — extrapolating from what it has learned to flag novel risks.
How It Integrates into a Development Workflow
DeepCode integrates directly with version control platforms. Developers connect it to their GitHub or Bitbucket accounts for cloud-hosted projects, and the tool reviews each commit automatically, flagging potential issues as part of the normal pull request flow. GitLab is supported for on-premises deployments.
For teams that want deeper integration, DeepCode also exposes an API, allowing larger enterprises to embed its review capabilities into custom internal tooling. At the time of this reporting (2019), the company was also working on its first IDE plugin, which would bring inline suggestions directly into the developer’s editor.
The workflow, in summary:
- Developer commits code to a connected repository.
- DeepCode reviews each commit automatically.
- Flagged issues are surfaced in the pull request interface or via API.
- Developers receive suggested fixes, not just error locations.
Claimed Accuracy and Time Savings
The obvious question for any automated review tool is: how accurate is it? Paskalev stated that developers typically spend around 30% of their time finding and fixing bugs, and that DeepCode can save roughly half of that — representing a meaningful reduction in overhead for active development teams.
The argument for the machine learning approach is scale: because DeepCode learns from the global open source community rather than any single team’s experience, it can in principle identify a broader range of issues than any individual reviewer or small group of reviewers. That said, no automated tool achieves perfect accuracy, and human review remains part of the process for edge cases and novel code patterns.
Background: ETH Zurich Spin-Out with Deep Research Roots
DeepCode was founded in 2016 as a spin-out from ETH Zurich (Swiss Federal Institute of Technology). The founding team includes CTO Veselin Raychev, formerly a software engineer at Google and a PhD researcher at ETH, and adviser Martin Vechev, a professor at ETH whose research group focuses on reliable and secure software systems.
That academic lineage gives the company a credible foundation in program analysis and formal methods — fields that underpin serious static and dynamic analysis tooling.
Funding, Pricing, and Language Support
In August 2019, DeepCode announced a $4 million seed round led by Earlybird, with participation from 3VC and Btov Partners. The company had previously raised $1 million in an earlier seed round.
Alongside the funding announcement, DeepCode revised its pricing model:
- Free for open source projects and educational use
- Free for enterprise teams of up to 30 developers
- $20 per developer per month for cloud deployments beyond the free tier
- $50 per developer per month for on-premises deployments
At the time, DeepCode supported Java, JavaScript, and Python, with plans to expand to C#, PHP, and C/C++ using the new funding. Known open source users of the platform included Embark Framework, the European Environment Agency (EEA), and PyMedusa.
Context: The Broader AI Code Review Market
DeepCode was operating in a market that was attracting significant investment. The automated software testing market was estimated at $8.5 billion in 2018, with projections for continued growth. Alongside testing tools such as Functionize, Testim, and Mabl, code analysis and review was emerging as a distinct and investable category.
The broader context included Microsoft’s $7.5 billion acquisition of GitHub and significant venture funding for GitLab — signals that the infrastructure supporting software development workflows had become strategically valuable at the enterprise level.
Note: DeepCode was subsequently acquired by Snyk in 2020. Readers researching this technology today should look into Snyk Code, which incorporates the DeepCode engine, for current capabilities and pricing.