Reminisce AI

Someone should be able to sit down, tell a family story naturally, and end up with something they can watch and keep. The hard part is everything between those two moments. People jump around in time, forget names, add context late, and care about details a model does not know are important. The product had to turn that into a coherent film without sanding away what made it personal.

I talked to more than thirty prospective users before the pipeline was mature: what they would want to preserve, who they would make it for, how comfortable they were recording themselves, and what would make the result feel worth keeping. That research changed the product several times.

Then I piloted it with twenty families. That is where the useful feedback came from: confusing onboarding, stories that generated badly, scenes that matched the transcript but felt emotionally wrong, and places the system needed more context than I expected.

I built the pipeline in Node.js and TypeScript. Raw speech becomes a transcript. From there the system has to understand people, places, events, relationships, chronology, and emotional beats, then turn that into a scene plan and structured inputs for visuals and video.

The hard part is coherence across the whole chain. A mistake early becomes a completely believable but wrong scene several steps later. Each stage has a clear output rather than one giant prompt deciding everything. I added retries and checks around generation because model calls fail in ways normal software does not: an API can succeed while the result is still unusable.

Technical correctness and product quality are different things. A scene can contain the right people and setting and still feel wrong to the family watching it. A generated film can complete successfully and still miss the point of the story.

That pushed the work toward evaluation, user intent, and where human judgment belongs inside the workflow — without treating the pipeline as a black box.