I Made a Fruit Fly Review My Code - Nick Nisi, WorkOS

Highlights

• A male fruit fly connectome containing 176,000 neurons, 6.3 million connections, and over 125 million synapses was used to review JavaScript pull requests.

• A system one model named Jev acts as the fruit fly nose by asking 30 unbiased questions about code, turning code diffs into literal smells.

• The fruit fly brain correctly identified code issues 88% of the time across general tests and up to 96% to 98% on harder tests.

• Team review skills and automated tools like Greptile and CodeRabbit can handle routine pull requests, escalating only risky code to human reviewers.

Timeline

1. 00:00 - 03:30

   The Code Review Bottleneck

   Software engineering has evolved past the point where writing code is the primary constraint. With advanced tooling handling code generation and automated checks, human reviewers are now the main bottleneck in shipping software. This shift calls into question the traditional purpose of human-led code reviews when functionality and automated validation take precedence.

   • Software development faces a major bottleneck because code is written faster and pull requests are more frequent than ever.
   • AI coding tools and continuous integration handle initial implementation and automated checks, leaving human review as the primary delay.
   • The historical definition of software bugs originated from a literal moth trapped in a computer relay in 1947.
2. 03:30 - 07:30

   Simulating a Fruit Fly Brain for Code Review

   To test unconventional review methods, a mapped fruit fly brain is integrated into the workflow. An AI system processes code changes and generates descriptive questions that simulate smells for the neural network. Depending on whether the code passes or fails tests, dopamine or bitter reinforcements train the model to register the code as acceptable or flawed.

   • A scanned male fruit fly connectome with 176,000 neurons and 125 million synapses serves as the core processing engine for code evaluation.
   • An AI model acts as the olfactory system by translating code diffs into non-biased questions that represent smells.
   • Good code releases sugar water smells and reward dopamine, while bad code produces bitter smells and punishment dopamine to guide memory retention.
3. 07:30 - 11:55

   Evaluating the Fruit Fly Reviewer

   Live demonstrations and testing confirm that the fruit fly connectome successfully differentiates between sound code and problematic code without human intervention. While it lacks the ability to explain why a change fails or suggest direct syntax improvements, its high accuracy rate highlights the viability of non-traditional review mechanisms.

   • Removing learned memories causes the fruit fly brain to malfunction, proving the biological connectome actively processes code quality.
   • Audience trials and automated testing showed the fruit fly achieved an accuracy rate between 88% and 98% in identifying problematic code changes.
   • The fruit fly reviewer cannot communicate specific line fixes or language context, relying instead on a binary reaction to code smells.
4. 11:55 - 16:31

   Redefining Human Roles in Code Review

   As automated agents and specialized models take over syntax and style checks, engineering teams can streamline their pipelines. By deploying team-maintained review scripts and risk-weighting classifiers, organizations can safely auto-merge low-risk pull requests. Humans retain oversight solely for high-risk changes, business logic alignment, and verifying proof of execution.

   • Human review should shift away from syntax checking toward verifying business logic, intent, and functional changes.
   • Teams can maintain specific review skills and automated classifiers to filter safe pull requests and only escalate high-risk changes to humans.
   • Verification practices like automated screenshots, video outputs, and smoke tests establish trust in fully automated merge pipelines.