Humanic Code

We kept running into the same problem while using AI to write code: getting a patch quickly is useful, but it is still hard to see what changed, what could break, and what deserves a closer human look. We wanted a focused review step between AI-generated code and acceptance.

=>What it does

Humanic Code is a local-first web app for reviewing AI-generated code changes before a developer accepts them. A developer can paste a unified diff or upload a patch, explain the expected behaviour, and review the change hunk by hunk.

The app scores each hunk as low, medium, or high risk. It pays particular attention to authentication, validation, database queries, dependencies, configuration, payments, and tests. The reviewer can ask a focused AI question about a specific hunk, compare the change against the intended behaviour, and acknowledge every flagged risk before approval becomes available.

Approved reviews are stored locally in the browser with the complete diff so the same audit can be reopened later. Humanic Code does not connect to Git hosting providers or perform remote merges. The developer stays responsible for the final decision.

=> How we built it

We built Humanic Code as a Next.js web application with a unified-diff parser, scoped risk heuristics, syntax-aware review cards, and IndexedDB for browser-local audit history. The local-first design avoids server-side file writes and works on serverless deployments.

We used Codex throughout the project to shape the review workflow, implement and refine risk scoring, improve the user interface, add local-first history, test the complete review-to-history path, and prepare the deployment. GPT-5.6 was part of our Codex collaboration for the core development work documented in the repository. The optional OpenAI review path uses the Responses API for hunk-level findings and questions when an API key is configured. The app is still testable without a paid key through clearly labelled deterministic fallbacks.

=> Challenges we ran into

The hard part was making the app helpful without pretending that a risk score can replace a developer. We avoided a remote merge button and instead built an acknowledgement gate, an intent-to-diff scope check, and a local audit trail. We also had to make the experience deploy safely on Vercel, where writing to a server file system is not reliable.

=> Accomplishments that we are proud of

We are proud that Humanic Code is a working, testable review tool rather than a static demo. It supports a full flow from real diff input to risk review, hunk explanation, explicit acknowledgement, approval, and private re-audit history. We also kept the app usable for judges without a paid API key.

=> What we learned

We learned that the most useful AI developer tools make the reasoning around a change more visible. The best output is not just a score; it is a clear workflow that helps a developer understand a change and decide responsibly.

=> What's next for Humanic Code

Next, we want to add configurable team review policies, richer language-aware analysis, and optional repository integrations while keeping the approval decision with the human reviewer.

Built With

• indexeddb
• monaco-editor
• next.js
• openai-api
• react
• tailwind-css
• typescript