An event-driven AI code reviewer for Uber's internal review platform that generates, grades, filters, deduplicates, and posts findings while leaving engineers in control of the reviewed change.
Submitted diff on Uber's review platform to posted, rated review comments.
Ingest an eligible diff
When a developer submits a change, uReview filters out low-signal files such as configuration and generated code, then builds a structured prompt with surrounding code context[Source 1]
Specialized assistants
Standard, best-practices, and AppSec reviewers generate findings for different issue classes[Source 1]
Periodic benchmark evaluation; Claude 4 Sonnet as generator with o4-mini-high as grader was the highest-F1 reported pairing[Source 1]
Harness
A prompt-chained pipeline separates comment generation, confidence grading, validation, semantic deduplication, and category filtering; three specialized assistants were in operation when published[Source 1]
Sandbox
Not reported
Tool access
Reviews eligible code in Uber's six monorepos across Go, Java, Android, iOS, TypeScript, and Python; richer internal artifacts were not yet connected when published[Source 1]Code-only access at publication; other internal artifacts are named as not yet connected.
Knowledge
Surrounding source context plus a shared registry of Uber-specific coding and style rules[Source 1]
Context management
Comments and metadata are streamed through Kafka to Hive for feedback analysis, experiments, and operational dashboards[Source 1]The Kafka-to-Hive stream is feedback and operations context, not run-time context management.
Credentials
Not reported
Interfaces
Uber's Phabricator-based internal review platform and the CI process; GitHub is explicitly not the host.
Developer ratings, addressed-comment detection, and a curated benchmark tune prompts, thresholds, and models[Source 1]
Reported observations
These results come from the sources. Notes below each result explain what it covers and any reported limits.
Headline claim
Uber's introduction reports reviews of over 90% of approximately 65,000 weekly diffs, with over 75% usefulness and over 65% addressed comments; a later paragraph says 65,000 diffs per month[Source 1]
Conflicting evidence
Sources disagree about this statement. See the conflicting sources: 1.
Reported by
Uber
Scope
Introduction's reported weekly diff coverage, plus comment usefulness/addressed rates; later paragraph's monthly period conflicts and remains unresolved
Production coverage, developer ratings, and automatic addressed-comment detection
Observed in August 2025
About this result
Category
Adoption output
Evidence type
Reported measurement
What it covers
uReview coverage of Uber weekly diffs, usefulness-rated comments, and addressed comments, with the conflicting monthly figure unresolved
Key observation
Median review latency of 4 minutes across all six Uber monorepos[Source 1]
Reported by
Uber
Scope
Reviews across all six Uber monorepos
Method
Production latency telemetry
The source does not report the denominator of this figure.
Observed in August 2025
About this result
Category
Cost latency
Evidence type
Reported measurement
What it covers
Median uReview review latency across Uber's six monorepos
Key observation
Approximately 1,500 developer hours reportedly saved per week, based on an assumed 10-minute second review per processed commit[Source 1]
Reported by
Uber
Scope
Processed commits excluding configuration files; modeled second-review time savings
Denominator
Over 10,000 commits per week
Method
Processed commits multiplied by an assumed 10 minutes for a second human review
Observed in August 2025
About this result
Category
Effectiveness
Evidence type
Estimate
What it covers
Modeled weekly developer hours saved by a second automated review, at an assumed 10 minutes per processed commit
Lessons
Reported opinion: Uber favors fewer useful review comments and suppresses categories that engineers consistently rate poorly.[Source 1]
uReview uses separate stages to generate, grade, filter, and consolidate comments, then records developer feedback.[Source 1]
Uber introduced uReview by team or assistant, tracking benchmark precision and recall alongside addressed comments and reported false positives.[Source 1]
Reported: The capture's pipeline order: ingestion and context, generation by specialized assistants, quality filtering, then in-line posting for developer ratings.
human involvement
Reported: Engineers remain in control of the reviewed change; usefulness ratings are optional feedback, not a required checkpoint.
implementation
Reported
validation
Reported: Addressed-comment detection re-runs uReview five times per final commit; a curated benchmark measures precision, recall, and F1 before release.
The opening paragraph reports approximately 65,000 weekly diffs, but the cost discussion says 65,000 per month. These conflicting periods remain unresolved; neither is independently verified.
Reported by
Uber
Scope
Introduction's reported weekly diff coverage, plus comment usefulness/addressed rates; later paragraph's monthly period conflicts and remains unresolved
Introduction reports reviews of over 90% of approximately 65,000 weekly diffs; cost discussion instead says 65,000 diffs per month
Duplicate representation of Headline claim. Component of the compound headline: the same over-90% coverage of approximately 65,000 weekly diffs, with the same unresolved monthly conflict.
Statement type
Metric
Provenance
Reported
Confidence
Low
Confidence reason
The opening paragraph reports approximately 65,000 weekly diffs, but the cost discussion says 65,000 per month. These conflicting periods remain unresolved; neither is independently verified.
Reported by
Uber
Scope
Introduction's reported weekly diffs analyzed; monthly period in the cost paragraph conflicts and remains unresolved
Denominator
Approximately 65,000 weekly diffs according to the introduction; same volume described as monthly in the cost paragraph
Over 65% of posted comments addressed in the same changeset
Duplicate representation of Headline claim. Component of the compound headline: the same over-65% addressed rate over posted comments.
Statement type
Metric
Provenance
Reported
Confidence
Medium
Confidence reason
Uber estimates addressed comments by rerunning generators and matching comments by semantic similarity. This supports an estimated response rate, not confirmation that each issue was fixed.
Reported by
Uber
Scope
Posted comments considered addressed in the same changeset
Denominator
Posted comments
Method
Five reruns on the final commit and semantic-similarity matching
The source estimates about 1,500 hours using over 10,000 commits and a 10-minute assumption; the rounded figures do not arithmetically reconcile exactly and are not observed time savings.
Reported by
Uber
Scope
Processed commits excluding configuration files; modeled second-review time savings
Denominator
Over 10,000 commits per week
Method
Processed commits multiplied by an assumed 10 minutes for a second human review
The precision section describes confidence ratings, category pruning, and duplicate removal. The preference is specific to uReview's reported feedback.
