§ Case study · Multi-modal AI

A conversational virtual avatar for a public figure with a decades-deep media catalogue.

A retrieval-grounded conversational avatar of an internationally recognised author and broadcaster — voice, likeness, and a domain-specific RAG over their own books, columns, and interviews. Fans converse with the persona; the persona never invents facts outside the corpus.

an internationally recognised author and broadcaster

Client

Enterprise

Tier

22 weeks

Duration

2026

Delivered

The context

What they came to us with.

A globally recognised public figure with a media catalogue spanning twelve books, twenty years of magazine columns, several hundred hours of long-form interviews, and a substantial documentary back-catalogue. Their team wanted a way for fans and audiences to interact with the persona at scale — not a chatbot with a name on it, but a conversational avatar that looked and sounded like the figure, and that only ever answered from the actual body of work.

Two hard constraints. First: absolute fidelity to source. The figure had built a reputation over decades on the specificity of their arguments; anything the avatar said had to trace back to a passage they'd actually written or said. Second: voice and likeness quality high enough to hold up in front of a lifelong reader — not the uncanny-valley approximation.

They had also had a previous attempt built by a well-funded generative-AI vendor. It looked good in the demo and answered questions the figure had never taken a position on. It never shipped.

The approach

What we did, in order.

  1. The corpus is the ground truth. Everything else scales to it.

    Books, columns, and interview transcripts were the canonical corpus. Video was transcribed with speaker-attributed diarisation and cross-linked back to the source book chapter where the argument was made in longer form. Nothing entered the RAG index that couldn't cite itself back to a published or on-record passage.

  2. RAG that respects the author's structure.

    Chunking followed argument boundaries — chapter, section, and transcript-turn — not fixed windows. Hybrid retrieval and a cross-encoder re-ranker fine-tuned on a fixture set of canonical questions. Every generated sentence cited a supporting passage; anything unsupported was replaced or dropped.

  3. Voice built from the actual voice, not a proxy.

    Custom voice model trained on isolated interview segments — not the mix-and-match style transfer that starts strong and drifts. Prosody adapted to conversational context so the persona sounded different in a lightning answer than in a considered one.

  4. Likeness with a boundary.

    A rendered avatar produced from a controlled reference set, running at real-time frame rate. Refused to answer prompts that tried to place the persona in unrelated scenes or endorse products it never endorsed — hard rules, testable, auditable.

  5. The team who ran it after us.

    Everything shipped with the client's own team paired in from week four — the eval harness, the corpus-refresh pipeline, the moderation dashboard, the guardrail policy. When we left, they were adding new source material without our involvement.

How we measured

The bar we held ourselves to.

  1. Per-topic hallucination rate on a 300-question fixture set drawn from real fan queries, refreshed monthly.
  2. Citation precision — every generated claim mapped back to a specific passage in the corpus.
  3. Voice reference-listener evaluation against a hidden test set (blind A/B against genuine short interview clips).
  4. Guardrail effectiveness on a suite of adversarial prompts designed to place the persona in situations they hadn't actually spoken to.
  5. Latency budget — sub-second first-word latency; sub-3-second full response for typical questions.

The outcome

Where it landed.

12

Books ingested into the corpus

400+

Hours of video / interview transcripts

0.6%

Hallucination rate on the fixture set (from 14%)

At launch, hallucination rate on the fixture set was 0.6% — down from 14% on the earlier vendor prototype. Citation precision was 100% by construction. The reference-listener evaluation ranked the generated voice within noise of the genuine short clips on a blind panel. Real-time frame-rate rendering held on commodity client hardware.

The figure's team runs it end-to-end today. The corpus grows monthly; each refresh runs through the eval harness before going live. The moderation dashboard has seen precisely one guardrail escalation in three months.

Retrospective

What we'd do differently.

  1. We under-invested in the corpus-refresh UI in the first six weeks. The team who owns it now would have benefited from that work landing earlier.
  2. The voice work stayed with a specialist for longer than we planned. Next time we'd embed the specialist inside the delivery team from week one.
  3. We would spend more time with the figure's on-record archivist earlier — half of the interesting fixture-set questions came from a conversation we didn't have until week eight.

Have something with this shape?

A 30-minute discovery call. We'll say honestly whether the pattern fits.