A closed Q&A pilot built for AI analysis, live by Tuesday

The constraints

What made this hard?

Three things shaped the build. The first was speed. "Half a day of focused work, live by Tuesday" is an ambitious brief for a web app with a Supabase back-end and a bespoke attribution mechanism. It meant no scope creep, nothing exploratory, every decision made fast.

The second was the anonymity/attribution paradox at the centre of the project. Respondents shouldn't have to identify themselves — that changes how honestly people answer. But the founder needed the data attributed to named individuals so the reports were useful. Any mechanism that asked respondents to enter their name defeats the purpose; any mechanism that left the data unattributed defeats the usefulness. The solution had to live somewhere the respondent never sees.

The third was tone. The founder's main commercial brand is fast, bright, construction-adjacent. This project needed to feel like the opposite: quiet, considered, nothing urgent. That distinction had to be built in from the start, not applied as a visual layer afterwards.

The approach

How did Tincture frame the problem?

I built the questionnaire app on Vercel with a Supabase back-end. The attribution mechanism lives entirely on the admin side: the founder enters a participant name, and the system generates a unique URL. When that URL is shared and a response is submitted via it, the response lands in Supabase tagged to the participant name — without the respondent ever seeing the name, entering anything identifying, or knowing their link is unique. The URL carries the attribution; the respondent doesn't.

Every structural decision in the build was made with AI processing in mind. Responses land in Supabase in a clean, typed structure — no free-text noise, no inconsistent formatting, nothing that requires cleaning before an AI model can read it. The founder can pull a participant's full response set and pass it straight to a model with a persona prompt (therapist, career coach, life auditor) and get a coherent report back. The app doesn't do the analysis; it makes the analysis possible without any preparation work between capture and AI input.

The build

What was shipped?

The questionnaire content was supplied by the founder; I built the capture layer, the unique-link generator, the admin interface, and ran a QA pass before go-live. The design direction was deliberately calm — muted palette, generous whitespace, slow pacing, nothing that signalled urgency or demanded attention.

The pilot workflow: the founder shares unique links with an initial group of around six contacts, who can forward their link onwards to their own people. Each submission arrives in the store tagged to the participant the founder originally named, regardless of how many hops the link has travelled.

The outcome

What were the results?

Pilot live by Tuesday, as agreed. Unique-link generation working end-to-end: the founder creates a participant record in the admin interface and gets a clean, shareable URL back in seconds. Responses landing tagged in Supabase, structured and ready for manual AI processing per participant. No email capture from respondents, no accounts, no follow-up flow, privacy completely preserved -exactly as scoped.

The pilot is now running. The core mechanic- anonymous response collection with silent participant attribution - is working as designed.

The takeaway

What's the transferable principle?

If you're building for AI analysis downstream, the data capture layer is the design problem worth solving first. The AI report is only as good as the structure of what it reads, and "we'll clean it up before we run it through the model" is a tax you'll pay on every single participant. Building the structure in from the start — typed fields, clean schema, no free-text noise — means the AI step is copy, paste, prompt. The anonymity mechanism matters too, and the URL-level solution handles it cleanly: the respondent feels anonymous because they are anonymous, and the researcher gets attributable data because the URL already carried the name.