AI Readiness & Diagnostic Platform

A full-stack diagnostic platform that processes an 11-question user intake and returns a structured readiness roadmap scored across six weighted dimensions by Claude Sonnet.

Every result is held in a pending state in Supabase until a human consultant approves it in the admin dashboard — satisfying EU AI Act Article 14 mandatory human oversight before the result reaches the user.

Standard diagnostic tools either deliver instant unreviewed model outputs or require fully manual consultant processing for every case — neither scales responsibly.

The design challenge was automating the scoring while preserving a structurally enforced human approval gate at delivery.

A user completes the 11-question intake. Claude Sonnet scores the response across six weighted dimensions and generates an overall readiness score, dimension breakdown, and month-by-month roadmap as structured JSON.

The result is stored as pending in Supabase and an n8n webhook fires a notification to the consultant admin dashboard. The consultant reviews and approves before Resend delivers the result email to the user.

FastAPI handles intake processing, Claude Sonnet scoring via the Anthropic SDK, and Stripe Checkout for the optional deliverable kit. Supabase (EU-Frankfurt) stores all diagnostic results. LangSmith (EU endpoint) traces every model call for observability.

• FastAPI with slowapi rate limiting on all scoring endpoints.
• LangSmith EU endpoint (eu.api.smith.langchain.com) for full model call tracing.
• Supabase Frankfurt for GDPR-aligned data residency.

No diagnostic result reaches a user without explicit consultant approval. The pending state in Supabase is a data constraint, not a policy — the Resend delivery step cannot fire without an approved flag.

• Results stored as pending until consultant sign-off in the admin dashboard.
• EU AI Act Article 14 human oversight satisfied by system architecture, not process.
• LangSmith traces available for every scoring decision.

The scoring model produces an overall readiness score (0–100), a breakdown across six weighted dimensions, and a structured month-by-month roadmap as a single JSON output.

LangSmith observability allows review of every Claude call, output structure, and any schema deviations.

The platform is live and actively iterated. Current focus areas are structured output consistency, deliverable kit expansion, and load testing against the rate limiting layer.

• Intake, scoring, human review, and Stripe kit purchase are all in production.
• slowapi rate limiting and LangSmith observability are in place for production scale.
• Structured output schema validation is the active reliability improvement.

The most consequential design decision was where to place the human gate — after scoring but before delivery, enforced at the data layer.

That constraint shaped the entire system: the Supabase schema, the n8n notification logic, the admin dashboard, and the Resend trigger all follow from the decision that no model output reaches a user without explicit human review.