AI/ML · Full Stack
ReviewPilot
Multi-tenant GenAI code-review SaaS
- Status
- Complete
- Year
- 2026
- Live application
- Runs locally with Docker Compose. Billing is a mock; not publicly deployed.

- deterministic review rules
- 17
- deterministic review rules
- tenant roles with server-side enforcement
- 4
- tenant roles with server-side enforcement
- automated tests across layers
- 123
- automated tests across layers
01 Problem
AI code review is only useful if findings point at real code. ReviewPilot runs 17 deterministic rules first, then a grounded model pass whose findings must quote the exact diff lines they reference. Usage is metered against plan limits in a multi-tenant SaaS shell.
02 Product overview
A NestJS API with session auth and tenant guards accepts diffs by paste, upload or HMAC-signed webhook. A BullMQ worker runs the review pipeline (parse, static rules, model, validate, persist) and streams progress over SSE. Findings can be triaged, commented on, compared across versions and exported to JSON, CSV, Markdown or SARIF 2.1.0.
Implemented
- Organizations with Owner, Admin, Reviewer and Viewer roles; email-bound invitations; audit log
- Cross-tenant access returns 404, never 403
- 17 deterministic rules followed by a grounded model pass, streamed over SSE
- Triage (accept, dismiss, false positive, reopen), comments, versions and compare
- Export to JSON, CSV, Markdown and SARIF 2.1.0
- Plan limits, usage metering, org API tokens, idempotent submissions and retention
Planned, not yet built
- MFA / SSO
- Real billing provider
- Production mail transport
- Git-host write-back
- Cloud deployment
03 System architecture
04 Core technologies
Backend
- NestJS
- Prisma
- PostgreSQL 16
- BullMQ
- Redis 7
- Zod
- argon2id
AI
- Ollama (qwen3:4b-instruct)
- Grounding validator
Frontend
- Next.js 16
- React 19
- TanStack Query
- Tailwind 4
- Plotly
Quality
- Vitest
- Playwright
- GitHub Actions
- gitleaks
05 Engineering challenges
A validator that was too strict
The first grounding validator rejected all five suggestions from the model. After fixing how quotes were matched, four were kept and one was correctly rejected.
Failing safely
If the model is unavailable, reviews degrade to rule-only results instead of failing.
Tenant isolation
Cross-tenant requests return 404 so resource existence is never leaked.
06 Technical decisions
- 01Server-side sessions instead of JWTs
- 02Rules first, model second, and the model must quote the diff
- 03Custom unified-diff viewer instead of Monaco
- 04A labelled fixture LLM for deterministic tests
07 Verified test results
- Backend unit
- 43
- Backend integration (real Postgres + Redis)
- 63
- Frontend unit
- 8
- Playwright
- 9
Latest GitHub Actions run on main: passing. Counts as reported in the repository's README and testing docs.
08 Screenshots
09 Links
No public deployment. Runs locally with Docker Compose. Billing is a mock; not publicly deployed.


