Say2Form is Zetaver's own SaaS product. It lets an organisation create a working digital form three ways: describe it in text, describe it by voice, or upload an existing document and get the same form recreated digitally. Once a form exists, it can be filled by voice through a natural conversation, with every answer traceable back to the moment it was said.
The idea came from eleven years of building and running a US behavioral health EMR, where the biggest daily complaint was never the software. It was the paperwork. Clinicians finished their sessions and then stayed late typing notes and intake forms field by field. Say2Form was built to remove that step.
The first version of Say2Form was a single monolithic application with the AI wired directly into it. It worked as a demo. In real use it exposed three problems that a bigger prompt or a better model could not fix.
The AI could not be tuned where it mattered. Every form request, whether it came from a text prompt, a voice recording, or an uploaded document, went through one large model call. When output was wrong, there was no way to see which part of the process failed. We spent a long time trying to improve results through fine-tuning, and the gains were slow and inconsistent, because the real problem was not the model. It was that a single model call has no way to plan, no way to check what it needs, no way to check its own output, and no way to recover when it gets something wrong.
The system had no memory. Every doctor, every patient, every form started from zero. If a patient visited twice, the second visit knew nothing about the first. If a clinic had built two hundred intake forms, the two hundred and first was generated as though the clinic had never built one. Every session was a cold start, and the product never got better with use.
Sensitive data needed to stay away from the models. The product was born in healthcare. Patient names, dates of birth, medical details, and clinician notes all pass through it. Sending that information to a language model, or storing it in plain text, was not an option we were willing to ship.
Zetaver rebuilt Say2Form's AI layer from the ground up around three decisions.
Agentic workflows instead of one big prompt. We replaced the single model call with a graph of small, single-purpose agents, built on LangChain and LangGraph on top of a Django backend. All three form-creation pipelines (text prompt, voice, and document) now run through the same structure: one agent plans the form section by section, one gathers relevant context, one grades whether that context is enough to proceed, one generates the form, one validates it against structural rules, and one repairs it if the validation fails. If the request is too vague to act on, the pipeline stops and asks the user instead of guessing.
This is what fixed the tuning problem. Instead of trying to train one model to understand every kind of request, each step could be inspected and improved on its own. We could finally see how a prompt was being analysed, how the response was being shaped, and where a form went wrong. A failure in one step no longer took the whole run down with it.
A memory architecture, so the platform learns from what it has seen. We introduced layered memory so that every form built, every correction made, and every conversation recorded leaves the platform slightly better at its job. The layers run from universal form-design knowledge, down through industry patterns, the organisation's own history of forms, the way a specific professional works, and finally what is known about a specific patient or client. When two layers disagree, a fixed order of authority decides which one wins, and validated clinical instruments always sit at the top so nothing can reword them.
The practical effect: a returning patient no longer starts from scratch. Known details are pre-filled and the professional confirms rather than re-enters. A clinician's own abbreviations and phrasings are recognised. A clinic's forms start looking like that clinic's forms. The system knows what it already knows.
Privacy built into the pipeline, not bolted on. Personal information never travels to an AI model as-is. Before any text reaches a model, direct identifiers such as names, dates of birth, phone numbers, emails, and record numbers are replaced with placeholders, and the real values are restored only after the result comes back. The detection that finds those identifiers runs locally, so the data being protected is not sent offsite in order to protect it. Form submissions and patient records are encrypted at rest, so stored sensitive data is never held in plain readable form. Every organisation runs in its own isolated database, not a shared table, so one customer's data is structurally separated from every other customer's.
Say2Form is live and demo-ready, and we are open about where it stands: it is at the stage of onboarding its first users and design partners, so this case study reports what the rebuild changed in the product rather than customer metrics.
What the rebuild delivered:
Form generation from text, voice, and documents runs through inspectable, self-correcting pipelines instead of a single opaque model call, which turned a slow fine-tuning struggle into a process the team can tune step by step.
The platform now retains and applies what it has learned. Returning patients, familiar professionals, and an organisation's existing forms all shape the next output instead of being forgotten between sessions.
Sensitive data is shielded before AI processing and encrypted in storage, which is what allows the product to be used in healthcare settings at all.
Every AI-generated form and every voice-filled submission still goes through a human review before anything is saved or submitted. The AI drafts; the professional decides.
Most AI form tools are a prompt box in front of a model. That works in a demo and breaks in a clinic. Say2Form is built the way production AI systems have to be built: work broken into agents that can be checked and corrected, a memory that makes the product better with use instead of identical every time, and a privacy layer that keeps patient and client data away from the models entirely. It is the same discipline Zetaver brings to client AI work, applied to our own product first.
Contact us to learn more about how we can build a similar solution for your business.
Get in Touch