Relay — AI Engineering Command Center

A Codex and GPT-5.6 command center that analyzes complete repositories, scores engineering layers, finds risks, proposes refactors, and independently verifies fixes in an isolated sandbox.

Comment

Inspiration

Engineering teams often use separate tools to understand unfamiliar code, investigate incidents, review risks, propose changes, and verify fixes. AI can accelerate these tasks, but recommendations alone are not enough. Teams need source evidence, cost controls, reproducible verification, and a clear human approval boundary.

Relay was created to turn that fragmented workflow into one auditable AI engineering command center.

What it does

Relay analyzes public GitHub repositories in read-only mode. It inventories the complete repository tree, divides eligible files into bounded domain batches, and produces:

  • An architecture map
  • Engineering ratings across eight layers
  • Evidence-backed findings with file and line references
  • Refactoring opportunities
  • An executive engineering summary
  • A prioritized and safe change plan
  • Incident root-cause and blast-radius investigation
  • Before-and-after fix verification in an isolated sandbox

Every analysis is identified by repository, branch, commit SHA, and batch content. Unchanged results can be retrieved from cache, preventing unnecessary model calls and reducing cost.

The flagship incident workflow accepts a production signal, traces the likely root cause, proposes a regression test and minimal patch, and executes the before-and-after proof in an isolated Vercel Sandbox.

Human approval remains explicit, and Relay never modifies the analyzed repository.

How we built it

Relay uses Next.js, React, and TypeScript for the product experience; GitHub APIs for read-only repository ingestion; deterministic analysis modules for architecture, scorecards, heuristics, and refactoring checks; GPT-5.6 through the OpenAI Responses API for semantic review and incident reasoning; Vercel Runtime Cache for commit-aware reuse; and Vercel Sandbox for isolated verification.

Semantic results are structured with paths, lines, confidence, evidence, and remediation. Relay validates the evidence against the source supplied to the model before displaying it.

If the OpenAI API is unavailable, deterministic heuristic analysis remains available and is clearly identified as a fallback.

How we used Codex

Codex collaborated throughout repository inspection, product design, implementation, debugging, testing, cost-control development, deployment, browser verification, and documentation.

I made the final product and safety decisions, including read-only access, evidence validation, human approval, isolated verification, and the distinction between complete eligible-file coverage and an impossible guarantee of zero defects.

Codex Session ID: 019f7f38-e542-7751-86c6-d9a062389c3d

Challenges we ran into

  • Grounding semantic findings in exact source evidence
  • Keeping API keys and credentials strictly server-side
  • Avoiding repeated AI cost for unchanged commits
  • Covering large repositories without an arbitrary global file limit
  • Distinguishing generated fixes from independently executed fixes
  • Reproducing before-and-after behavior in an isolated environment
  • Presenting a complex engineering workflow clearly and responsibly

Accomplishments that we're proud of

  • Complete eligible-file repository coverage with explicit exclusions and failures
  • Structured GPT-5.6 semantic review with heuristic fallback
  • Commit-aware cost estimation and cache
  • Engineering scorecard across eight layers
  • Evidence-based refactoring signals
  • Evidence-linked findings and safe remediation plans
  • Incident-to-fix workflow with isolated before-and-after execution
  • Explicit human-control and read-only safety boundaries
  • A public, responsive, credential-free judging experience

What we learned

AI engineering tools become more useful when they communicate uncertainty and levels of proof. A model finding, a generated patch, a passing test, and an independently verified fix are different forms of evidence and should never be presented as equivalent.

We also learned that cost and context management belong in the product experience. Commit identity, bounded context, caching, and coverage reporting are user-facing trust features, not merely infrastructure details.

What's next for Relay

  • Cross-domain dependency analysis beyond direct import cycles
  • Historical engineering-score trends across releases
  • Approval-gated GitHub integration for explicitly authorized repositories
  • Team collaboration and policy-based approval workflows

Built With

Share this project:

Updates