Tom & I, an oral-history voice agent
A browser-based voice interviewer for oral histories, built for the Tom of Finland Foundation: a persona designed to listen rather than steer, a six-stage topic arc and tiered consent captured as structured data.
- Company
- Independent · Tom of Finland Foundation
- Role
- AI & Digital Transformation Consultant, Nonprofits (independent); designed and built
- Timeline
- 2026
- Platforms
- Web
- Type
- AI
TL;DR
- Problem
- An oral history depends on an interviewer who listens and on consent handled properly. I wanted to see whether a voice agent could do both for a small nonprofit without trapping it with vendors.
- What I did
- I designed and built Tom & I for the Tom of Finland Foundation: a browser-based voice agent on ElevenLabs, with Antigravity and a Notion data layer, an interviewer persona designed to listen rather than steer, a six-stage topic arc and T1/T2/T3 tiered consent captured as structured data.
- Result
- Development is paused.
Context & problem
Tom & I was an oral-history voice agent I built for the Tom of Finland Foundation, where I’m Board Director & Secretary and Technology Committee Chair, as part of my independent AI work with nonprofits.
An oral history depends on an interviewer who listens, and on consent handled with care. The question I set out to answer was whether a voice agent could do both, without locking a small nonprofit into one vendor.
Discovery & insight
I started from the interviewer rather than the technology. In an oral history the person telling the story should lead, so the persona, Aino, was designed to listen rather than steer.
Strategy & bets
I compared four voice stacks, ElevenLabs, OpenAI Realtime, Vapi and Retell; the build was on ElevenLabs. The larger bet was a principle: every component swappable and cost-checked, with self-monitoring agents, so a small nonprofit isn’t trapped by vendors. I costed it against a planning assumption of 200 interviews at 20 minutes each.
Solution
Tom & I was browser-based, on ElevenLabs, with Antigravity and a Notion data layer. The conversation was designed as a six-stage topic arc, built as a conversational state machine. Consent was captured at session close as structured data in three tiers, T1, T2 and T3.
Execution
I designed and built it myself, from the persona and the conversation design to the consent capture and the cost plan.
Validation & results
Development is paused. The signals I would have watched were whether a participant keeps talking without being led, whether consent lands as structured data at session close, and what an interview costs against the plan.
Reflection
What I’d keep from Tom & I is the principle more than the stack: for a small nonprofit, being able to swap a component and see what it costs matters more than any one vendor’s feature. I’d also keep treating consent as part of the conversation design rather than paperwork added after it.