From a capped no-code prototype to a production AI wellness app

HeyFern is an AI somatic therapy app that coaches people through nervous-system regulation in real time. We designed and built it end to end, and migrated the entire live product off the Base44 no-code builder onto an owned full-stack TypeScript codebase, with zero data loss and zero downtime for existing users.

For an early-stage founder, the takeaway is simple: the product went from a no-code ceiling to a billable, safe, scalable app without a months-long rebuild that burned runway and churned users.

The challenge: AI you can safely put in front of vulnerable users

Most AI products can absorb a bad answer. HeyFern cannot. It speaks to people who are already activated, so an off-tone or unsafe model response is not a UX annoyance, it is a real risk. At the same time, the live app and its users were locked inside Base44, with no clean path to add the safety, reliability, and billing infrastructure a real wellness product needs.

Two jobs, then: get off no-code without losing a user, and make the AI safe enough to ship to people who are not okay.

What we built

A full-stack TypeScript product. React and Vite front end, a Hono API on Node, PostgreSQL with Prisma across 18 models, Better Auth (email and Google), Stripe subscriptions, and AWS S3. 28-plus pages and 63-plus REST endpoints.

A real-time activation check-in. Users report where their nervous system sits on a 1 to 10 scale, and Claude guides them through regulation from there, adapting as that number moves. Voice input and text-to-speech let a session be spoken with eyes closed, which is the point when the goal is to down-regulate rather than stare at a screen.

A custom AI safety layer. This is the real engineering. Every model response passes through an output-guard system with seven issue-type detectors before a user ever sees it. If a response trips a detector, a retry loop of up to six attempts with corrective instructions kicks in, so the user gets a safe, on-method reply instead of a raw model output. Telemetry logs every check and auto-prunes at 5,000 events, and an admin dashboard lets the team tune safety policy live without a redeploy. Daily limits are enforced atomically with serializable transaction isolation, so they hold under concurrent load instead of leaking through race conditions.

A clean migration off no-code. A custom ETL pipeline moved the entire production app off Base44 to an owned stack with zero data loss. Existing users carried over seamlessly; the team gained a codebase they can actually extend. It is the same playbook we run for MVPs stuck on Lovable, Bolt, v0, Cursor, Manus, or Figma Make: keep the data and the users, lose the ceiling.

Why founders care

The slow killer for an early-stage SaaS is time to value. Owning the stack unlocked Stripe billing, an admin control surface, and an AI safety system no no-code tool could deliver, all while the live app kept running. And because the model layer is provider-agnostic, the underlying LLM can be swapped or upgraded as the field moves, with no re-architecture and no vendor lock-in.

The result

HeyFern ships today as a production AI wellness app at heyfern.app: full-stack TypeScript, real subscription billing, a seven-detector AI safety layer with a self-correcting retry loop, and a no-code migration that kept every existing user. A real product in a domain where getting the AI wrong was never an option.