02 / Representative code-review demo

Repair and improve existing code

Make inherited or AI-assisted code safer, clearer, and easier to extend.

Deterministic local samples. No model call or private code is involved.

Before and after code panes side by side with a repaired arrow between them and three green pass badges: request token per load, body read before the guard, stale response blocked. — Before / after repair

Concept illustration of a magnifying glass and wrench repairing a cracked code block until a check badge appears. — Concept illustration: inspect, repair, verify.

Problem

AI-assisted code can be fast to produce and expensive to trust. The risk is rarely the syntax alone: it is the missing boundary around inputs, stale async state, authorization assumptions, duplicated logic, or an error path that disappears into a catch block.

Approach

I have worked with AI-generated and AI-assisted code by reviewing it, fixing defects, and refactoring the result into code that can be reasoned about. This portfolio demo uses representative samples so the method is public without exposing a client codebase or inventing a client outcome.

Proof

Each sample has a concrete defect, a repaired version, and deterministic checks that explain why the repair matters. The viewer can switch samples and run the checks without an API key.

Limits

The samples are intentionally small teaching cases. They demonstrate review judgment, not a measured customer result, model benchmark, or claim about every generated codebase.

Product walkthrough

What it looks like in use

Real states of the interactive demo below: what a reviewer sees when the flow succeeds and when it fails.

Screenshot of the code repair desk demo showing the repaired async state guard sample and its deterministic checks passing. — Repair desk / after state with checks passingThe interactive demo's repaired state: real code text, checks run locally.

Interactive / Representative code-review demo

Repair desk: inspect a defect and run the checks

Deterministic local samples. No model call or private code is involved.

Selected representative sample

Async state guard

For detailed review

Engineering diagrams

Denser diagrams of the same boundaries, for reviewers who want the full structure.

Diagram of three recurring risks in generated code: a missing async state guard, a caller-controlled input boundary, and a validation error path hidden by a catch block, each with its repair. — Three risk pointsAsync state, input boundary, validation error: the risks reviewed first, with the repair pattern for each.

Continue the review

Bring the existing codebase, constraint, or failure path.

This page demonstrates a bounded working method. The next step would be to inspect the actual repository and agree on the smallest verifiable slice.

Open the evidence page