PocketBooks SaaS MVP

The problem

PocketBooks was a bootstrapped SaaS idea targeting small service businesses — freelancers, consultants, and small agencies — who were managing their accounting in spreadsheets or paying for QuickBooks features they didn't need. The founders had validated the concept with 50 potential customers but had no product and limited runway.

Their key differentiation: AI-powered transaction categorization and monthly reconciliation. Not just bookkeeping software, but bookkeeping software that actually does most of the work. They needed an MVP fast enough to capture the momentum from their validation, and a technical partner who understood both SaaS architecture and how to integrate AI into a product experience properly.

What we built

We scoped the MVP in week one: transaction import (bank and card), AI-powered automatic categorization, monthly reconciliation with confidence scoring, invoice creation and tracking, and a simple dashboard. Auth, Stripe billing, and a customer onboarding flow. Nothing beyond what was needed to test whether users would pay.

The AI reconciliation feature — the core differentiator — uses a fine-tuned classification model trained on common small business transaction types, with merchant name normalization and a confidence threshold that flags low-confidence categorizations for user review. First-time categorizations improve the model for future similar transactions.

The entire frontend is Next.js. Backend Node.js on Supabase. Plaid integration for bank connections. Stripe for billing. Deployed on Vercel in week eight.

Technology

• Next.js + TypeScript (frontend + API routes)
• Supabase (PostgreSQL + auth)
• Plaid API (bank data integration)
• OpenAI fine-tuned model (transaction classification)
• Stripe (subscriptions + billing)
• Vercel deployment

Results

Launched with 12 paying beta customers at the end of week eight. Three months post-launch: 87 paying subscribers, 94% transaction auto-categorization accuracy, average monthly close time for users under 20 minutes versus the 3+ hours they reported doing it manually. The founders are now raising a seed round with the product as the primary proof point.