Case study

Stash

Everyone saves things on Instagram and never goes back. Stash pulls every save, watches the video, listens to the audio and reads the caption, then turns the pile into categorised action items and step-by-step tutorials you can actually use.

Client: Stash
Year: 2026
Role: Product design · Full stack · AI
Stack: Next.js · AI Gateway

0 saves · 0 processed · 0 in the engine

Your saves, finally doing something.

saves pulled

content watched & read

open action items

step-by-step tutorials

The graveyard → action

847saves sat untouched

0you ever came back to

∞good ideas, lost

Auto-sorted

263

AI & Agents

Coding & Dev

224

Design & UI

Productivity

140

Content

Making & 3D

107

Business

Fitness

The problem

A save is a promise to your future self that you almost never keep. The content is good, the intent is real, and then it goes into a grid of thumbnails you will never scroll back through. The value is locked inside video and audio, which is exactly the format you cannot search, skim or act on.

What I built

A pipeline that treats each save as a document to be understood rather than a thumbnail to be stored. It transcribes the audio, reads the on-screen text and the caption, then decides what the save actually is: a thing to do, a technique to learn, or noise. The output is a categorised library, not a feed.

Watch, listen, read

video, audio and caption are all processed, because the useful part is rarely in the caption

Action or tutorial

every save resolves to something you can do or something you can learn, or it is dropped

Re-sortable

categories are proposed by the model and can be re-run, so the shape of the library is not frozen at import

Watch it work

One save, through the engine

queued

saved 62 days ago

Building a floating oak shelf

Stash is watching

Stash is listening

Stash is reading

What came out

A save nobody was going back to.

Sorted by meaning

Categories are not folders the user maintains. They are proposed from what the saves are actually about, and the whole library can be re-sorted when that shape stops fitting. The work of filing is the work nobody does, so the product does it.

The honest bit

This is a personal product with one user. There are no adoption numbers here because there are none to report. What it shows is a messy input, video and audio nobody wants to sit through, turned into something structured enough to act on.

Got a pile of content nobody has time to go through?

Tell me the job you want to stop doing