
Your Digital Human Twin Should Be Able to Show Its Work
Published: November 18, 2026
Someone asked me what a digital human twin is good for. I sent a link, because the honest answer needs about 3 minutes of talking to a simulation of a person and one click on the sources under its reply. Everything before that click sounds like marketing.
Most people picture a chatbot wearing a personality. What I build now looks more like a search index with a mouth. A pile of dated fragments pulled from one person's own record, an index that finds the right fragment for a question and a voice that speaks only from what that index returns.
I run this in production at regenesis.chat, where the roster is made of people whose entire public record survives: proclamations, patents, lectures, letters, essays. The rest of this post is what it took to get the thing from a convincing demo to something I trust enough to ship.
The Demo Passes in 5 Minutes
The first prototype was a strong model, a persona prompt and a folder of sample quotes. It was excellent for 2 minutes. Then the third question arrived and the twin answered in a voice that belonged to nobody, with facts it had read somewhere else. The voice still sounded right, and nothing underneath it could be checked against a source.
That gap is the whole problem in this category. Voice cloning and lip-sync are solved enough that a 20-minute job produces something people believe. Believing is the danger. A twin that improvises in someone's accent does more damage than one that refuses to answer, and the audience cannot tell which one they are watching.
So the product had to stop being a performance and start being a record.
Material, Index, Voice
Three layers, and only the third one is visible.
Material. One document is one file with a date and a name. Letters, transcripts, lectures, books, court statements. Public only, nothing from private correspondence that was never published. The date matters as much as the text: it decides whether an answer belongs to 1803 or to the exile years.
Index. Two retrieval passes over the same query, a dense one for meaning and a sparse one for exact wording, plus a date filter. What comes back is a set of fragments instead of a summary. Summaries are where invention enters.
Voice. The corpus gets spoken live, with video, audio and a transcript beside it. This layer is the cheapest to build and the one people judge the product on, which is a trap worth naming.
The first two layers decide whether the twin has anything true to say. The third decides whether anyone listens.
No Trace, No Answer
The rule I hold the build to is simple and expensive to implement: every sentence carries an address.
Each reply ships with the fragments behind it, the file names, the dates and the fragment offsets. On the live roster that trace is now the proof block under the answer, something like 5 fragments, the newest dated 1821 and 3 named sources. If the material does not support a sentence, the sentence does not appear.
The best example of this came from a question I did not expect the twin to handle well: when did you realise the Russian campaign was lost? The answer I got was a refusal in the person's own register. It said, in effect, that the retreat made the outcome clear in November, but that there is no date and no line in the papers where the author himself writes that down, and that the memoir written years later explains the decision instead of recording it.
That answer is worth more than any fluent paragraph. It is also the part a generic model gets wrong by default, because a generic model always has something to say.
What the Live Build Looks Like
Numbers from the site as of 10 November 2026, not from a pitch deck:
- 10+ personas live, 46 more in the queue
- 565 sources in the index
- 92,790 fragments retrieved from
- 190 ms search time
- about 1 second to the first spoken word once the call connects
Search speed turned out to matter more than I expected. Latency above roughly 2 seconds reads as hesitation, and hesitation breaks the illusion of a person who is thinking. Getting the index warm put the visible pause back into a range where the conversation feels natural.
The Voice Is the Part People Trust Too Much
Once the lip-sync is right, people stop reading. They ask follow-up questions, they take notes, they quote it. That is exactly when the label matters, so the label is on screen for the entire conversation and every answer is signed as a clone.
Three rules shipped with the product, and none of them are decoration:
- Labelled. Simulation, not person. On every answer, in the UI, with no way to switch it off.
- Public only. Published interviews, books, speeches, letters. Nothing scraped from private messages or leaked archives.
- Removable. One request from the person or their estate pulls the material out of the index.
For historical figures these rules cost almost nothing. For a living person they decide whether the product is legitimate at all, which is why the roster starts where the record is finished and public.
Where a Twin Earns Its Keep
The commercial version of this is boring and that is why it works. An institution with 40 years of documents and a phone queue of the same 30 questions. A founder whose answers live in their head and nowhere else. An onboarding flow where the person you would ask has no time. A museum, an archive, a publisher sitting on a back catalogue.
In each of those cases the value sits in the corpus, and the twin is a way to make the corpus talk. Which also means the buyer is not paying for the avatar. The avatar is the cheapest line item in the invoice.
5 Questions Before You Buy a Twin
I get asked to review builds that other people shipped, so here is the list I use.
- How many dated sources, and what is the newest date? A count without a horizon hides a stale index.
- What happens when the answer is not in the corpus? If the twin invents, the rest of the numbers do not matter.
- Can I see the fragment, or only a summary? The summary is the model's opinion about the source.
- Can a person or estate pull their material, and how fast? An answer of "we will look into it" is a legal problem waiting.
- Is the label on screen while it speaks? Not in the terms, on the screen.
A vendor who cannot answer question 2 in one sentence has built a costume.
The Record Is the Product
Digital twins of people are not a model problem. The models are good enough to hold up a conversation, and they were good enough a year ago. The hard part is building a private library that is clean, dated, searchable and honest about its own gaps, then refusing to let the model speak outside it.
That refusal is the feature customers end up paying for, even when they arrive asking for the voice.
If you want to see the difference between a performance and a record, the free tier gives you 3 minutes with any persona on the roster and the full source trace under every answer. Talk to one of them at regenesis.chat, then ask it something its papers cannot answer.
Sources
- regenesis.chat product roster, corpus counts and pricing, checked 18.11.2026
- AI Product Engineering: What I Learned Shipping AI Products People Actually Use on the earlier build
- Stop Writing Prompts. Start Writing Skills. for the persona pack approach behind the answering layer
Building with AI beyond this article?
I run The Agentic Architect Lab, live builds, agent workflows, and a playbook for technical founders shipping solo. No toy demos.
Join the Lab