ReviewFlow AI

The Problem Code review is one of the most valuable and most time-consuming parts of a software team's workflow. Reviewers must simultaneously check for security vulnerabilities, missing tests, broken API contracts, and code quality issues — all while reading unfamiliar code under time pressure. The result is inconsistent coverage: a reviewer who is strong on security may miss test gaps; one who focuses on logic may overlook a missing authorization check. Junior reviewers lack the pattern recognition to catch what senior reviewers spot immediately. And when the PR queue is long, review depth shrinks. The Solution ReviewFlow AI gives every pull request a structured, evidence-backed first-pass review before a human reviewer opens the diff. It reads the team's own code-review guidelines, inspects the Git diff between branches, and runs four parallel focused review agents — Security, Tests, Impact, and Code Quality — each scoped to exactly one concern. The results are consolidated into a short reviewer brief: a prioritized list of findings with file and line references, missing test cases, downstream impact analysis, reviewer questions to investigate, and a final recommendation (Approve / Approve with comments / Request changes). The tool never auto-approves or merges. Every output surface carries the statement: "AI findings are suggestions; human review remains required." The AI accelerates the reviewer — it does not replace them. Target Users Software development teams of any size who conduct pull request reviews. The primary user is the human code reviewer: ReviewFlow AI hands them a structured brief so they spend their attention on verification and judgment rather than first-pass scanning. Secondary users are developers who want early feedback on their own branches before requesting review. How They Interact With It A reviewer runs npm run review from the terminal, or opens the web dashboard at https://review-flow-ai-r1f8.vercel.app/ . The dashboard displays the risk level, all findings with severity badges and file references, missing tests, impact notes, and the reviewer questions the AI identified. The reviewer validates each finding against the diff, answers the questions, and makes the final call. What Makes It Creative and Unique Most review tools run static linters or single-pass analysis. ReviewFlow AI uses IBM Bob 2.0 in Agent mode as the orchestration layer: Bob reads the team's guidelines document, spawns four role-isolated subagents in parallel, and consolidates their output into a single brief. The subagent prompts in review-engine/prompts/ enforce evidence-first output — every finding must cite a concrete file and line, and must distinguish confirmed evidence from a suggestion. This makes the output fast to validate rather than slow to second-guess. The architecture is also transparent and extensible: the four review functions map directly to the four Bob subagents, and each can be upgraded to a live watsonx.ai inference call without changing the report structure or the dashboard. In short: ReviewFlow AI compresses the first pass of a pull request review from minutes of unfocused reading into a structured brief a senior reviewer can validate in under two minutes — consistently, on every PR, against the team's own rules.