Danijel builds Back to work

Offhand

See what Instagram knows about you.

01

Overview

A personal Android app that turns your own Instagram data export into a report on your habits.

Its AI runs on the phone: the classification and the profile happen on the device, and captions never leave it.

Product
Android app
Role
Solo: product, design and build
Focus
On-device AI, privacy
Tech stack
Kotlin · Jetpack Compose · ONNX Runtime
Built with
Claude Code
The Mirror Report opening screen Trend and projection screen What they fed you screen Feed diet screen AI profile screen
Screens rendered from the app’s code with sample data
02

What I owned

It is my “built with AI” project: I designed and built it on my own with Claude Code.

  • The product. What the report shows, and how to turn a habit into something you can see.
  • The design. Every screen, from the opening report to the AI profile.
  • The build. A native Android app in Kotlin and Jetpack Compose, built with Claude Code.
  • The on-device AI. A two-stage classifier and a bundled multilingual model running through ONNX Runtime.
03

The problem

Instagram learns your habits from everything you do, but you never get to see that picture yourself.

You can download your own data, but it arrives as raw files that nobody can read. The evidence is there; the story is not.

04

The solution

A report on your own habits

Import your data export, and the app turns it into a report you can scroll through.

Trend and projection screen

Trend and projection

Monthly activity with a 12-month projection if nothing changes, so a habit becomes a direction you can see.

What they fed you screen

What they fed you

What you did next to what the feed served back: likes, saves and comments against posts shown, videos watched and ads served.

Feed diet screen

Feed diet

Every post is classified on the phone, so you can see what your feed is actually made of.

AI profile screen

AI profile

On-device AI turns those numbers into a personality profile, with evidence for each trait.

Cheap first, model second

Classifying the feed uses two stages. A fast keyword pass handles most posts. Only what it can’t place goes to a small multilingual model, because the content mixes English with other languages and keywords don’t carry across languages. Matching the tool to the task: the model only runs where the keywords run out.

Private by design

The phone had no built-in model I could use, so I bundled my own: a compact multilingual model, quantised to run through ONNX Runtime on the device. The classification and the profile both run on the phone, and captions never leave it.

05

Outcome

You can see where your attention goes on Instagram: the time, the habits over time, what the feed served you and a profile of yourself, with the analysis running on your own phone.

On-deviceClassification and profile run on the phone
MultilingualA bundled model for the posts keywords can’t place
SoloDesigned and built by me with Claude Code

Status: a working app, built as a personal project.