A builder, not a prompt box
The brief started as free-text prompting. I pushed it toward a structured builder instead: niche, gender presentation, skin tone, hair, body type, outfit, personality, and voice tone, each a real choice.
01 · The pain
What was actually broken.
- Becoming an influencer usually means your real face and real voice are online forever, no undo once it's out.
- Most AI persona tools give you one generic face and call it done. Nobody lets you actually choose her hair, skin tone, body type, and voice like a character, not a filter.
- Every text-to-image tool generates a new stranger on each regeneration. There's no way to keep her the same person across a hundred future posts.
02 · Who this is for
Persona.
A solo creator or small brand building an online presence Wants a consistent character to post as, fashion, fitness, gaming, whatever the niche, without putting their own face, body, or voice online permanently. Doesn't want a new random stranger every time they generate an image either. Wants her to stay the same person.
03 · The problem
What the design had to solve.
Build a tool where constructing a virtual persona is a real product decision, chosen attributes, not a sentence you have to know how to write, and then actually give her content to post: a photorealistic portrait, a written voice, a spoken voice, and a lip-synced video.
07 · Process
From skeleton to final screens.
Three AI systems chained into one persona
NVIDIA's flux.1-dev renders the portrait, Llama 3.1 writes her name, bio, and a sample caption in character, and Cartesia turns any script you type into her actual speaking voice, picked from a real voice catalog I mapped to gender and tone by hand.
Made the photo actually talk
Wired in D-ID for lip-synced video, then hit a real bug: their own image-upload endpoint hands back an internal reference their pipeline silently never processes, jobs sit forever, zero credits charged. Isolated the exact cause with a control test against their own demo image, then routed around it by hosting the generated portrait myself.
Shipped live, then chased what broke
Deployed to Netlify from day one instead of building a local mockup. Found and fixed a function timeout killing image generation mid-request and a CSS specificity bug that kept a loading spinner stuck over finished images, both caught by actually driving the live site on a real phone viewport, not just reading the code.
09 · Goals & intended outcomes
What this was designed to do.
- Prove a persona can be genuinely designed from chosen attributes, not randomly generated by a prompt, and still look photorealistic.
- Give the persona a real voice, not just a picture: written voice, spoken voice, and a talking video.
- Ship a working version fast and verify every claim against the live, deployed app instead of trusting that code should work.
- A live, working builder: pick her attributes, get back a photorealistic portrait, a name and bio, a spoken voice reading any script you write, and a lip-synced talking video.
- Found and fixed three real production bugs during the build: a Netlify function timeout, a silently broken third-party upload pipeline, and a CSS specificity bug hiding a loading state, each confirmed with real data, not assumptions.
- Verified mobile responsiveness on a real iPhone viewport and caught a UI bug there before calling the feature done.