Quality assurance · Production

uReview: Uber's multi-stage GenAI platform autonomously reviews over 90% of 65,000 weekly code diffs, saving approximately 1,500 developer hours per week

The problem

As code volume grew—amplified by AI-assisted development—Uber's reviewers became overloaded and struggled to catch subtle bugs, security vulnerabilities, and best-practice violations consistently, leading to missed errors, slower feedback loops, production incidents, and slower release cycles.

First attempt

Third-party AI code review tools required GitHub hosting, but Uber uses Phabricator; those tools also produced many false positives and low-value true positives, and could not integrate with Uber's internal systems. Simple standalone LLM prompts generated too many false-positive comments that eroded developer trust.

Workflow diagram · grounded in source
1
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.”
2
File filtering and context building
ai_action
“It filters out low-signal targets such as configuration files, generated code, and experimental directories. For the remaining files, the system builds a structured prompt that includes surrounding code context such as nearby functions, …”
3
Specialized assistant comment generation
ai_action
“- The Standard Assistant detects bugs, incorrect exception handling, or logic flaws. - The Best Practices Assistant enforces Uber-specific coding conventions by referencing a shared registry of style rules. - The AppSec Assistant targets…”
4
Multi-layered quality filtering
validation
“uReview runs them through a multi-layered filtering process: 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 …”
5
Comment posted on review platform
output
“The system posts validated comments directly on the code review platform, in line with the code.”
6
Developer feedback collection and 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…”
7
Automated addressed-comment 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, with 75% of its comments rated as useful and over 65% addressed in the same changeset—outperforming human reviewers whose comments are addressed only 51% of the time—saving approximately 1,500 developer hours weekly, equivalent to nearly 39 developer years annually.

Reported metrics
Weekly diffs analyzed (coverage rate)over 90%
weekly diffs at Uber~65,000
Comments rated as useful by engineers75%
Posted comments addressed in same changesetover 65%
Show all 13 reported metrics
weekly diffs analyzed (coverage rate)over 90%
weekly diffs at Uber~65,000
comments rated as useful by engineers75%
posted comments addressed in same changesetover 65%
human-written comments addressed as bugs in same changeset51%
commits processed weeklyover 10,000
estimated human reviewer time per commit (benchmark)10 minutes per commit
developer hours saved weeklyapproximately 1,500 hours
developer years saved annuallynearly 39 developer years annually
CI review median latencymedian of 4 minutes
monorepos coveredall six
AI cost vs third-party toolsorder of magnitude less
sustained usefulness rate targetabove 75%
Reported stack
Claude-4-Sonneto4-mini-highApache HiveApache KafkaPhabricator
Source
https://www.uber.com/en-GB/blog/ureview/?uclick_id=0a73d271-32e7-4b77-9697-a587a4c8d9fe
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, with 75% of its comments rated as useful and over 65% addressed in the same changeset—outperforming human reviewers whose comments are addressed only 51% of th…

What tools did this team use?

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

What results were reported?

Weekly diffs analyzed (coverage rate): over 90%; weekly diffs at Uber: ~65,000; Comments rated as useful by engineers: 75%; Posted comments addressed in same changeset: over 65% (source-reported, not independently verified).

What failed first in this deployment?

Third-party AI code review tools required GitHub hosting, but Uber uses Phabricator; those tools also produced many false positives and low-value true positives, and could not integrate with Uber's internal systems.

How is this quality assurance AI workflow structured?

Developer submits change → File filtering and context building → Specialized assistant comment generation → Multi-layered quality filtering → Comment posted on review platform → Developer feedback collection and streaming → Automated addressed-comment evaluation.

Related quality assurance cases