Projects / Built · Case study
Persona architecture for creator AI clones
A platform that lets creators and influencers build an AI version of themselves from their face, voice and personal background, and hold conversations with fans. My part: the design that keeps every creator’s clone consistent, and remembering what matters about them.
- Context
- Creator-economy platform
- My role
- Onboarding design, clone pipeline, prompt architecture, memory, evaluation
- Stack
- Azure OpenAI, Soul Machines, Deepgram, ElevenLabs, Langfuse
- Focus
- Agent memory and where it stops scaling
Two design decisions
A creator cannot hold a one-on-one conversation with every fan. A clone that carries on authentic conversations at scale has to behave the same way for every creator, and remember the creator’s life across many separate conversations.
- One template, many creators. A single standardized prompt template, with creator-specific content injected dynamically, keeps every profile consistent without a hand-written prompt per creator. Changing persona behavior becomes a template edit, not a rewrite.
- Memory outside the context window. I built the memory stores that let a clone recall a creator’s life events reliably across conversations, not only within one session.
What else I did
- Designed the onboarding flow, starting from what creators wanted reflected in their avatar, to make collecting face, voice and personality inputs easy.
- Built and optimized the automated clone-creation pipeline, and integrated a specialist avatar vendor for the face and voice, instead of building that multimodal capability in-house.
- Validated the real-time streaming response pipeline end to end through rigorous testing.
- Built an LLM-as-a-judge evaluation framework for conversational quality and persona fidelity.
The scaling limit I identified
The current design injects a creator’s full backstory into every prompt. That works, and it will not scale as backstories grow. I explored a lorebook-style alternative, the technique roleplay platforms use:
Every fact goes into every prompt, whatever the fan asks.
A fact is surfaced only when the conversation triggers it. Example: a fan asks about “your last trip”.
Illustration with made-up example facts, not real creator data.
| Full backstory in every prompt | Lorebook-style injection | |
|---|---|---|
| How it works | Everything about the creator goes into each prompt. | A background fact is surfaced only when conversation keywords trigger it. |
| Strength | Simple; nothing relevant is missed. | Scales with large backstories. |
| Status | Current design | Explored, not built |
Limits
- I report no fidelity or quality results here, because none are in my records, and no speed, scale or usage figures.
- Handling of uncensored creator personas is still being evaluated.