Quality assurance · Production

uReview: Scalable, Trustworthy GenAI for Code Review at Uber

The problem

Uber's code reviewers were overwhelmed by increasing code volume from AI-assisted development, with limited time to identify subtle bugs, security issues, or consistently enforce best practices — leading to missed errors, slower feedback loops, production incidents, wasted resources, and slow release cycles.

First attempt

Third-party AI code review tools required GitHub (Uber uses Phabricator), generated many false positives and low-value suggestions, could not interact with Uber's internal systems, and cost an order of magnitude more than the internally built solution.

Workflow diagram · grounded in source

Developer submits change

trigger

“When a developer submits a change on Uber's code review platform, uReview first determines which files are eligible for automated review.”

File eligibility filtering

validation

“It filters out low-signal targets such as configuration files, generated code, and experimental directories.”

Structured prompt construction

ai_action

“the system builds a structured prompt that includes surrounding code context such as nearby functions, class definitions, and import statements. This context helps the language model produce precise and relevant suggestions.”

Specialized assistant comment generation

ai_action

“uReview uses a pluggable assistant framework, where each assistant focuses on a specific class of issues. This pluggable framework allows each assistant to be developed and evaluated independently, and use customized prompts and context.…”

Multi-layered quality filtering

validation

“A secondary prompt evaluates each comment's quality and assigns a confidence score. The prompt is customized for each assistant type, and confidence thresholds for pruning are set at a fine-grained level (per assistant, per language, and…”

Comment posted inline

output

“The system posts validated comments directly on the code review platform, in line with the code.”

Developer feedback and data streaming

feedback_loop

“Developers can rate each comment as "Useful" or "Not Useful" and optionally add a note. All comments, along with their associated metadata—including assistant origin, category, confidence score, and developer feedback—are streamed to Apa…”

Automated re-evaluation

feedback_loop

“It automatically evaluates if a given posted comment has been addressed by re-running uReview five times on the final commit. Because LLM is stochastic, a single rerun might skip a lingering issue or revive one that is already fixed, so …”

Reported outcome

uReview analyzes over 90% of Uber's weekly ~65,000 diffs, maintains a sustained usefulness rate above 75%, saves approximately 1,500 developer hours per week (nearly 39 developer years annually), and delivers feedback within a median of 4 minutes per commit across all six monorepos.

Reported metrics

Weekly diffs analyzedover 90%

weekly diff volume at Uber~65,000 diffs per week

Comment usefulness rate75%

comments addressed rate (uReview)over 65%

Show all 11 reported metrics

weekly diffs analyzedover 90%

weekly diff volume at Uber~65,000 diffs per week

comment usefulness rate75%

comments addressed rate (uReview)over 65%

comments addressed rate (human reviewers, for comparison)51%

developer hours saved weeklyapproximately 1,500 hours saved weekly

developer years saved annuallynearly 39 developer years annually

CI review turnaround timemedian of 4 minutes

commits processed per weekover 10,000 commits

human reviewer equivalent time per commit10 minutes per commit

cost vs third-party toolsan order of magnitude less than what typical third-party tools charge

Reported stack

uReviewClaude-4-Sonneto4-mini-highApache HiveApache KafkaPhabricator

Source

https://www.uber.com/en-IN/blog/ureview

Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

uReview analyzes over 90% of Uber's weekly ~65,000 diffs, maintains a sustained usefulness rate above 75%, saves approximately 1,500 developer hours per week (nearly 39 developer years annually), and delivers feedback…

What tools did this team use?

uReview, Claude-4-Sonnet, o4-mini-high, Apache Hive, Apache Kafka, Phabricator.

What results were reported?

Weekly diffs analyzed: over 90%; weekly diff volume at Uber: ~65,000 diffs per week; Comment usefulness rate: 75%; comments addressed rate (uReview): over 65% (source-reported, not independently verified).

What failed first in this deployment?

How is this quality assurance AI workflow structured?

Developer submits change → File eligibility filtering → Structured prompt construction → Specialized assistant comment generation → Multi-layered quality filtering → Comment posted inline → Developer feedback and data streaming → Automated re-evaluation.