@gamechangerai: A company wanted families to store the memories of a person and later engage with them through a conversational 3D avatar built from photographs and recordings. The avatar is the visible part. What I actually spent the time on was the constraints this subject matter puts on the architecture. I had to design for two distinct people. One is alive and recording their own memories calmly over months. ⠀ The other has just lost someone and is opening the application in grief. The same data store serves both, and almost nothing else about their sessions should match. Not the pacing, not the tone, not how much the system offers before it is asked. Memory collection runs as an interview rather than a form, so the tool asks a question, listens, and recurses into what the person just said. No single tool turns photographs into a talking 3D likeness, so I compared several published face reconstruction and motion transfer models, chose one, modified its code and retrained it. ⠀ Voice synthesis and speech to text came from managed cloud services. The part most teams skip is that I built against published work on the stages of grief. It changed what the avatar volunteers and when the system suggests stopping. A system that speaks in the voice of a dead parent can do real harm, so I treated the psychological model as a safety requirement. Would you talk to one? ⠀ Follow @gamechangerai for more. ⠀ #ai #machinelearning #generativeai #digitaltwin #aiethics

Rahul Kumar

Open In TikTok:

Region: PT

Monday 21 September 2026 18:03:07 GMT