Join the alpha waitlist
Be first to ship AI code you can trust.
We respect your privacy. Unsubscribe anytime.
Built on 500+ enforcement events · 8,000+ knowledge surfaces · 228 real-world rules
The AI quality problem
Every major study says the same thing: AI-assisted development breaks quality in predictable ways. Your linter can't see it. Your tests can't catch it. By the time it shows up in production, the debt has compounded for months.
Every 25% increase in AI adoption correlates with a 7.2% decrease in delivery stability — second year running.
DORA, State of DevOps 2024
AI-generated code gets reverted at double the rate of human code. The itinerant-contributor pattern at scale.
GitClear, 153M-line study
Roughly 1 in 3 AI-generated code samples contains a security vulnerability. Most pass review.
Academic & industry consensus
Tools like Backstage and Cortex measure your org.
PillarCI measures what your AI is doing right now, and makes it fix it.
How it works
PillarCI sits between your AI assistant and your repo. It watches every edit, checks it against your project's rules, and blocks what would quietly break things — before the commit.
A machine-readable scorecard for your codebase's maturity. Four tiers: Foundation → Product → Production → Scale. Boolean capability gates, weighted quality metrics, and abstraction-quality multipliers keep projects from pretending infrastructure sprawl is maturity.
Assessments emit structured JSON prescriptions with fix commands, validation steps, and routing hints for concurrent AI instances. Hooks block writes that would leave the codebase worse than it started. No more "it'll be fine" workarounds.
shadcn-style ownership for backend. Auth, caching, queues, logging — all ship with clean adapter interfaces and two-layer ownership (base + user). Swap SQLite → Postgres or in-memory → Redis without touching your services.
Live dogfood
These are real numbers from our own production codebase (PersonaMind, ~150K LOC TypeScript monorepo). No cherry-picked demos — just 30 days of the system catching what would have shipped without it.
Source: yarn health · 30-day window · metrics continuously emitted to JSONL streams
The Four Pillars
A project on SQLite with clean adapter boundaries is more mature than one on Postgres + Redis + S3 glued together with raw queries. PillarCI measures abstraction quality, not technology choice.
Full guardrail suite, externalized config, repository pattern. Coverage >40%, complexity ≤15.
Auth, structured logging, errors. Interface-segregated DB access. Coverage >70%, zero critical findings.
Rate limits, caching, queues, APM. Dialect-agnostic migrations. Coverage >80%, code health >7.
Feature flags, circuit breakers, versioning, SLOs. Tested infra swaps. DORA: daily deploys, <1hr lead time.
Why PillarCI
Backstage has scorecards. shadcn has drop-in components. Cursor has AI. None of them close the loop: assess → prescribe → enforce.
| Capability | Backstage | Cortex | shadcn | PillarCI |
|---|---|---|---|---|
| Maturity scorecards | ✓ | ✓ | — | ✓ |
| Drop-in components | — | — | ✓ | ✓ |
| Machine-readable prescriptions | — | — | — | ✓ |
| Hook-level enforcement (blocks AI) | — | — | — | ✓ |
| Multi-instance AI orchestration | — | — | — | ✓ |
| Abstraction-quality over infra-count | — | — | — | ✓ |
Roadmap
Keyword & file-path aware knowledge surfacing. 228 rules in production. Trigger-condition DSL with block/warn/context severities.
yarn tier --json emits structured prescriptions with fix commands, routing hints, and validation. Drives the pre-task loop.
Auth-JWT, caching, queues, logging — each with two-layer ownership and adapter interfaces. shadcn CLI flow for NestJS + Kysely + tRPC.
File-lock registry + prescription queue. Route prescriptions across concurrent AI agents. Natural coalescing, safe parallel edits.
Early-access users get first look at the alpha, a private roadmap channel, and founder-level response times.
No spam. No fluff. One email when the alpha opens.