The Thin Line Between Vibe Coding vs. AI-Assisted Engineering
The Context: Bridging the Prototype-to-Production Gap
Our client, a venture-backed WealthTech platform, faced a common inflection point. Early development was characterized by extreme velocity—using AI to narrate design changes and ship features in hours. This vibe coding phase was successful in building a functional workbench for exit planning advisors, but as the platform began handling sensitive financial data, the limitations of ungoverned AI generation became clear.
We stepped in to transition the codebase from a collection of high-functioning prototypes into a cohesive, venture-ready infrastructure. The impact was not just in the code written, but in the systems established to ensure that speed no longer compromises institutional trust.
Engineering Impact: Hardening the Core
We moved beyond the fire and forget model of early AI assistance, implementing rigorous engineering standards across three critical domains:
- Complex State Management: The platform’s onboarding process involves multi-step, sequential data entry—a notorious source of intermittent state failures in client-side applications. We refactored this logic using Vercel Workflow DevKit, moving state management to the server. This ensures that advisor progress is deterministic and persistent, eliminating the phantom data loss that plagued the earlier reactive model.
- Proprietary RAG Pipeline and Content Integrity: We have been deeply involved in maturing the RAG pipeline that powers the platform’s core value: turning discovery data into actionable financial deliverables.
- Using FastAPI and LangChain, we built a pipeline that ingests client financials into a secure document store.
- We implemented LangSmith and Guardrails AI to create an evaluation layer. This ensures that AI-generated financial summaries are validated against safety guardrails before reaching the advisor, effectively neutralizing the risk of hallucinations.
- Infrastructure and Security Governance: The most significant impact has been the implementation of automated guardrails. Early in the transition, we caught a critical regression where an AI-assisted update nearly exposed the entire user table via a misconfigured Row-Level Security policy.
- We integrated Mesa.dev into the CI/CD pipeline. Specialized agents now automatically audit every PR for security leaks, FinOps cost spikes, and architectural alignment.
- We established a type-safe API layer using tRPC and Drizzle ORM, ensuring that the Next.js frontend and the Python backend remain in sync without the manual overhead of updating schemas.
The Strategic Result: Transitioning to Venture-Ready
By establishing these patterns, we have enabled the product owner to step away from the PR review bottleneck. The engineering team no longer just builds features; they manage a self-governing pipeline that produces enterprise-grade software at startup speeds.
We have successfully paid down the technical mortgage of the MVP phase. The platform is now structured as a monorepo, optimized for code discoverability, and protected by automated governance. This shift from vibe coding to AI-assisted engineering has provided the client with the deterministic security required to pursue their next round of institutional investment with confidence.
Tech Stack
- Frontend: Next.js (App Router), Tailwind CSS, tRPC, TanStack Query, ZooStand
- Backend: Fast API (Python), LangChain, LangSmith
- AI & Security: OpenAI, Anthropic, Groq, Guardrails AI, Mesa.dev
- Data & Infrastructure: Postgres (Neon), Drizzle ORM, Pinecone (Vector DB), Vercel (Deployment & Workflows)
- Authentication & Collaboration: Clerk, Live Blocks