Rescuing an AI-Generated Financial Modeling Application
Client: FinVector Analytics (Seed-Stage Fintech Startup) · Industry: Financial Technology
Auditing, refactoring, securing, and deploying an MVP built using Cursor and Bolt that was plagued with client-side secret leaks, infinite API loops, and broken auth.
Confidentiality Notice: Client branding and proprietary domain names are protected under mutual non-disclosure agreements. Architectural diagrams, code patterns, and performance metrics displayed below are authentic records extracted from the live production build.
The Problem
The founders used AI coding assistants to assemble a financial modeling prototype. While the visual charts looked impressive, the code had fatal flaws: Stripe secret keys were exposed in client bundles, database permissions allowed any user to read other tenants' financial files, and build scripts failed consistently on Vercel.
Key Constraints
- •Two weeks remaining before critical investor demo days
- •Zero tolerance for financial calculation inaccuracies or security leaks
- •Must preserve existing chart interaction UX that founders had carefully prototyped
Technical Approach & Execution
Conducted a forensic codebase audit, moved all sensitive operations to secure Next.js Server Actions, configured strict PostgreSQL Row-Level Security (RLS), and normalized the messy 60-file component tree.
Security & Secret Key Sanitization
Extracted all API keys and external queries from client React components into secure server-side routes with input validation via Zod.
Database Isolation with Row-Level Security
Re-engineered the Supabase PostgreSQL database schemas with strict RLS policies tied to validated JWT authentication claims.
Component Consolidation & Type Safety
Eliminated 35 duplicated component variations, resolved circular state updates, and achieved 100% strict TypeScript build passes.
Cloud & Deployment
- Next.js App Router on Vercel
- Supabase PostgreSQL with RLS
- Stripe Billing Integration
- GitHub Actions Automated CI Workflow
Tools & Languages
Key Engineering Lessons Learned:
- •AI code is great for generating UI ideas, but senior engineers are indispensable for auth, state, and security
- •Fixing the underlying architecture enabled the founder to continue using Cursor safely for small features
Explore other engineering engagements
See how our solutions scale across varying stacks and industries.
Real-Time Fleet Telemetry & Dispatch Platform
Rebuilding a fragmented legacy dispatch system into a reactive Next.js application with real-time GPS tracking, automated driver assignments, and automated invoice reconciliation.
WooCommerce to Headless Next.js & Sanity Migration
Transforming a sluggish monolithic WordPress/WooCommerce site into a headless architecture, reducing load times from 4.8s to 0.7s and raising mobile conversions.
HIPAA-Compliant Patient Telehealth & Scheduling Portal
Engineering an encrypted patient portal with integrated video consultation, instant appointment scheduling, and electronic health record integration.
Facing a similar AI & Vibe Coding Rescue challenge?
Let's review your architecture, codebase, or infrastructure constraints. You will speak directly with a senior systems engineer — not an account manager.