How I’d build Climb MVP 1

Client — mobile-first web app

Next.js (App Router) + TypeScript, installable PWA

A PWA ships day one to both platforms with no store review, and everything here is content and forms rather than device hardware. Native wrappers via Expo become worth it in MVP 2, when push notifications about a bureau response start driving retention.

Upload & storage

Direct-to-object-storage with short-lived signed URLs; private bucket, KMS-managed keys

The file never passes through an application server, which removes the single most common way credit reports leak. Source documents auto-expire after 90 days, and deletion is verified rather than queued.

Document extraction

Text-layer parse first, OCR fallback (AWS Textract or Google Document AI), then an LLM structuring pass over the extracted text

About 80% of uploads are bureau PDFs with a real text layer — parsing those directly is faster, cheaper and more accurate than OCR. OCR is the fallback, not the default. The LLM never sees a raw image and never invents a field: it maps extracted text to a strict schema, and anything under 90% confidence is confirmed by a human.

Report model

Postgres with a normalized tri-merge schema — report → bureau file → tradeline → derogatory → inquiry, versioned per pull

This is the heart of the product and the reason it can’t live in a spreadsheet-backed tool. Month-over-month comparison, cross-bureau conflict detection, and "what changed since last pull" all depend on a real relational model with history.

Analysis & guidance

Deterministic rules engine for what is disputable; LLM for explanation and letter drafting, grounded in the member’s own extracted data

The decision about whether an item is worth disputing is rule-based and auditable — you must be able to explain it to a regulator. The LLM writes the prose and answers questions; it never decides eligibility on its own.

Dispute & mailing

Letter versioning with member approval capture; print-and-mail API (Lob) behind an interface; daily USPS tracking poll

CROA makes the approval record legally meaningful, so the approved letter, its hash and the timestamp are stored together. The mail vendor sits behind an interface so it can be swapped without touching the dispute domain.

Subscriptions

Stripe Billing with metered mailing credits

Postage is a real per-unit cost. Modeling credits from day one keeps the unit economics visible instead of discovering them at scale.

Auth & identity

Managed auth (Clerk or Auth0) with mandatory MFA on any account holding a report

Rolling your own auth on a product holding consumer credit files is an unforced error. Managed auth also gives you SOC 2-relevant audit logging without building it.

Admin & analytics

Same Next.js app, role-gated routes; product analytics on internal IDs only

A separate admin app doubles the surface area for no benefit at this stage. What matters is that report access from the admin side is ticket-referenced and writes an immutable audit row.

Hosting

Vercel (app) + managed Postgres (Neon or RDS) + object storage in one region

Serverless is right for this traffic shape — bursty uploads, long-tail reads. The one caveat: use a pooled Postgres endpoint, or serverless connection storms will hang requests.

• Repo, CI, environments, error tracking
• Managed auth with MFA; role model (member / specialist / admin)
• Encrypted upload path with signed URLs and lifecycle expiry
• Data-handling decisions written down before any PII is stored

• Text-layer parser for the three bureau PDF formats
• OCR fallback path and page classification
• LLM structuring pass into a strict schema, with confidence scoring
• Human-review queue for anything under the confidence floor

• Versioned tri-merge schema and cross-bureau reconciliation
• Rules engine for disputability, with the reasoning surfaced to the member
• Score factor breakdown and the member-facing report screens

• Letter generation grounded in the member’s own data, with statute citations
• Review, edit and approval flow with an auditable approval record
• Mail vendor integration, certified mailing, tracking poll
• Response logging and the 30-day escalation trigger

• Coach grounded in the member’s report and an FCRA knowledge base
• Action plan and progress tracking
• Lesson library and moderated community with a pre-publication queue

• Stripe billing with metered mailing credits
• Admin console: members, mail queue, extraction queue, moderation, analytics
• Security review, penetration test, load test on the upload path
• Beta with 25–50 real members before public launch

CROA compliance is a product constraint, not a legal review at the end

The Credit Repair Organizations Act governs what you may charge for, when, and what you must disclose. It shapes the billing model and the approval flow directly — advance fees for dispute work are prohibited, and the three-day cancellation right has to be real in the UI. Design for it in week one.

Never promise a deletion

Every claim in the product has to be about process, not outcome. That constrains copy, the coach’s responses, and the marketing site. It is also the single clearest signal to a consumer that this product is different from the ones that got them here.

The bureaus are not an API

Report ingestion is document-driven because there is no consumer-permissioned API that gives you a full tri-merge file cheaply. Plan around documents. A data-provider integration is an MVP 2 conversation, and it changes the cost structure significantly.

Security review will happen whether you plan for it or not

Any payment processor, insurer or enterprise partner will ask. Budget the penetration test in week 13 rather than discovering it during a partnership conversation six months later.