Case Study — Jumbo Mana, Bonjour Vincent / Hello Vincent

What it is

Visitors speak into a microphone at an interactive terminal. On screen, a 3D Van Gogh sits on a stool in front of Wheatfield with Crows and answers in French or English. Answers are also transcribed on screen. The figure has a slight Dutch accent in some reports. The encounter is framed as a one-to-one conversation with the painter while he is painting, not as a docent standing in a lobby. S242 S243

Musée d’Orsay treated it as an addendum at the exit of a conventional exhibition, not as a replacement for paintings or wall text. That placement is part of the design: context first, conversation second. S242

Public names vary by language: Hello Vincent on the English Orsay page, Bonjour Vincent in French press and later Jumbo Mana marketing.

Who did it

This is a small deep-tech studio + one domain scholar + a national museum, not a Hollywood VFX house.

Pipeline

Jumbo Mana’s own site describes a four-stage delivery that matches what Orsay actually shipped. S220

1. Kickoff and editorial brief

• Character description, clothing, voice references, photos/paintings as likeness sources.
• Decision that the scene is Wheatfield with Crows — i.e. the last weeks at Auvers, the same frame as the exhibition.
• Hard question, owned by the museum: is this “the real Van Gogh,” a pedagogical tool, or a commercial stunt? Orsay insisted it had to be cultural, not only commercial. S248

Renaudineau’s public line: “We are able to bring these characters to life, but we are not trying to rebirth them.” S247

2. Corpus and “personality,” not Wikipedia

Primary fuel is Van Gogh’s correspondence. Counts in coverage:

• Musée d’Orsay: “some 900 letters.” S242
• Forbes: “more than 800 letters.” S134
• Press syndicates: ~900 letters plus early biographies. S247

Van der Veen’s job was to stop the model from “spitting back a Wikipedia page” and instead surface the painter’s own phrasing, reading list, emotions, and relationships. S243

Friess is explicit that humans still have to provision the data because “AI can say incorrect things.” S245

Sensitive-topic layer (necessary, not optional):

• The most asked visitor question was why he killed himself.
• Hundreds of visitors asked it.
• The team added human guidance so the model would steer suicide questions toward resilience rather than graphic invention.
• Answers still “constantly refine” by phrasing. Machine-learning errors about the life and work were reported in international press. S247 S250

That is the real pipeline risk of a generative histobot: the face can be locked and the letters ingested, and the model will still improvise.

3. Avatar, voice, behavior

• 3D sculpt of a recognizable Van Gogh (beard, blue eyes, missing ear) rather than a deepfake of a living actor. S134
• Behavioral / “agentic” AI is Jumbo Mana’s differentiator claim: not only text generation but social behavior, turn-taking, some physical autonomy. Treat the architecture claims as vendor language; the visible product is a talking 3D figure on a kiosk. S226
• Voice: slight Dutch accent reported; not a surviving recording of Van Gogh (none exists that could carry a conversation).
• Languages at Orsay: French and English.

This is closer to a game-engine character + LLM than to Dalí Lives (deepfake on an actor) or Dimensions in Testimony (unedited video of a real person).

4. Application and venue

• Microphone + terminal + on-screen avatar + live captions.
• Free, open access, after the paid exhibition path.
• Orsay framed visitor questions as training signal: “The more questions you ask, the more the AI learns and improves.” S242 That is a research stance, and it is also a governance problem (what gets written back into the character overnight?).

Renaudineau called the 2023–24 install an experiment in a research program that had started years earlier (one French radio piece says the program began “fin 2002,” which is likely a slip for the early 2020s). S243

5. Afterlife of the same character

The product thesis after Orsay: the same character can leave the gallery and live on a smart speaker or another museum. That is how a €0 visitor kiosk becomes a studio line.

Cost and personnel (what can be inferred)

No public invoice for Bonjour Vincent has surfaced. Reconstruct from company size and analogous builds:

• Studio side: a handful of AI engineers (EuroQuity: 6 of 13 in R&D) plus a 3D/avatar generalist and a product/UX person.
• Scholar side: one specialist historian, not a large curatorial committee.
• Museum side: digital-development lead plus exhibition curators who spent ~a year on the “is this him?” question.
• Infra: public supercomputer time (Jean-Zay) and regional/BPI grants — i.e. the French innovation system subsidized the first hero.

Ballpark relative to other histobots in the same research set:

• Cheaper than Dimensions in Testimony ($250k–$1.8M per living witness) because there is no light stage and no 2,000-question interview.
• More expensive than Hello History / Humy.ai (prompt + stock portrait) because of custom 3D, kiosk integration, and a named scholar.
• Same order of magnitude as a serious single-character museum agent: likely low-to-mid six figures all-in if you count grants, historian time, and museum staff, even if the startup invoice was smaller.

Personnel you actually need to copy the method:

1. Domain historian with veto power
2. Corpus editor (letters, not the open web)
3. Conversational designer for off-limits topics
4. 3D character + scene
5. Speech in/out in the venue’s languages
6. On-site integration and logging
7. A museum owner who will stand in front of the press when the model is wrong

Market success

What worked

• It shipped inside a destination exhibition at a top-tier museum. That is distribution most avatar startups never get.
• It created a reusable template (Bonjour [Name]) and at least two follow-on cultural characters.
• It opened a consumer channel (Alexa) and a B2B story (airports, retail, events — CMA CGM, Vinci, Galeries Lafayette, European Council appear on the studio’s own fundraising page). S229
• Press was global (Forbes, France Culture, franceinfo, LATAM syndications). For a ~12-person company that is the actual ROI of Vincent.

What did not, or is unproven

• No published visitor counts, dwell-time stats, or learning outcomes. Contrast Dalí Lives, which claimed 25,000 visitors in month one. S182
• Accuracy was contested in the same articles that celebrated it. A histobot that improvises will be fact-checked in public.
• Company finances remain early-stage ($30k disclosed, raising). Vincent is a reference customer, not a completed business.
• Suicide-as-top-query is a product warning: generative historical figures attract the most traumatic question first.

Commercial position vs peers

Vincent is the European museum-commission pattern: scholar-supervised generation, one hero per institution, then try to productize the engine.

Lessons for a replica build

1. Put the avatar after the real objects. Orsay’s exit placement is the ethical and pedagogical load-bearing wall.
2. Name the historian in the credits. Van der Veen is why this was taken seriously.
3. Budget a month for the one question you do not want asked. Here it was suicide.
4. Do not claim “no hallucination” if the model is generative. Jumbo Mana and Orsay were more honest than The Torch on that point; press still found errors.
5. Design the afterlife on day one (second venue, speaker, classroom) or the kiosk dies with the exhibition.
6. Keep the sentence Renaudineau used: bring the character to life, do not pretend to rebirth them.

Primary sources in the bibliography

S134 S220 S223 S225 S242 S243 S245 S247 S248

Full list: Citations — Historical Virtual Characters