companies

Diffblue

Oxford spin-out using reinforcement learning, not LLMs, to autonomously write and maintain Java unit tests inside the IDE and CI pipeline.

Diffblue was spun out of the University of Oxford in 2016 by Daniel Kroening and Peter Schrammel, both computer science researchers, on the premise that unit test generation could be automated more reliably with reinforcement learning than with hand-written scripts or generic LLM prompting. Mathew Lodge, a former VMware and Anaconda executive, joined as CEO in 2019 to lead its commercial push. Its flagship product, Diffblue Cover, analyzes Java code and writes JUnit or TestNG tests that compile, pass and add measurable coverage, generating tests the company says run roughly 250 times faster than a developer writing them by hand. Because it's trained with reinforcement learning against real compilation and test-execution feedback rather than predicting plausible-looking code, Diffblue argues its output is more reliably correct than what a general-purpose LLM produces for the same task. The company has raised more than $46 million total, including a Series A round led by Goldman Sachs and a 2024 extension backed by IP Group, AlbionVC and Citigroup, and counts Goldman Sachs, JPMorgan, Citi and S&P Global among its enterprise customers.

Founded

2016

Headquarters

Oxford, United Kingdom

Sector

applications