Paste a repo, and an agent tells you whether it will survive a deploy — with graph evidence instead of vibes. Correctness, not security: your guardrails stop the dangerous command; Preflight catches the broken one.
Built from scratch at HackwithBay 3.0 (2026-07-07, AWS Builder Loft SF) for the theme "Thoughtful Agents for Productivity."
- Sign in and point Preflight at a repository (Butterbase: auth, db, billing — deep reviews cost credits).
- The dashboard calls the review pipeline running as a managed endpoint on cloud.rocketride.ai.
- The pipeline builds a code property graph in Neo4j (imports, env reads, deploy targets, past findings) and derives every finding from a traversal: blast radius, load-bearing fan-in, prod env gaps, and past-failure proximity — see docs/graph-model.md.
- Deep reviews add a judge (an open model on Nebius AI Studio) and proof by execution: a Daytona sandbox clones the repo and actually runs its tests, attaching real output as evidence.
- The agent remembers: Cognee (open source, on the same Neo4j) recalls prior findings on re-review — and forgets the ones a human rejects.
| Piece | Tech |
|---|---|
| Backend, auth, payments, frontend hosting | Butterbase |
| Code property graph | Neo4j (Aura) |
| Review pipeline (managed endpoint) | RocketRide Cloud |
| Deep-review judge | Nebius AI Studio |
| Execution sandbox (bonus) | Daytona |
| Agent memory (bonus) | Cognee OSS |
docs/graph-model.md— the graph schema + the four Cypher traversals (committed first, before any pipeline code)pipeline/— the RocketRide review pipelineapp/— Butterbase functions + dashboardfixtures/demo-service— deterministic demo repo (missing prod env var, hardcoded secret, failing test, one load-bearing hub module)