I built an AI curation OS for Instagram saves between two projects. 246 installs. 84 MAU. Zero marketing.

Six months ago my Instagram saved folder had 500+ reels I'd never gone back to.

Recipes I meant to cook. Places I wanted to visit. Outfits I kept telling myself I'd wear.

The saving felt intentional. The forgetting didn't.

So I built Vestron. An app that AI categorises every reel you save and tells you who you are on Day 7 based on how you actually behave.

Here's everything honestly. What worked, what broke, and what I'd do differently.

The product

Vestron works through an Instagram bot. You DM it a reel. It gets categorised instantly — fashion, food, travel, tech, cinematic. Your vault organises itself while you do nothing.

You build a 7 day saving streak. On Day 7 Vestron assigns you one of 28 personality titles based purely on your saving behaviour. Not a quiz. Not a form. Your behaviour, observed.

"Keeper of Midnight" if you save after 11pm most nights. "Bringer of Storms" if you dump 15 things in 20 minutes. "Closer of Loops" if you actually act on what you save. "Collector of Intentions" if you commit to things and never do them.

We also shipped Wander. A community discovery feed showing what people with your aesthetic are saving globally.

The numbers

Total installs: 248 Monthly active users: 84 Daily active users: 6.6 Store conversion rate: 30.4% Firebase bill this month: ₹250 Marketing budget: ₹0

The 30.2% conversion is the number I'm most proud of. 1 in 3 people who find the Play Store page download it. The page is doing its job. The problem is almost nobody is finding it.

The night Meta broke everything

48 hours after V2 shipped, Meta deprecated the API we built our entire video experience on.

Old system: user DMs reel to our bot, webhook sends a CDN URL, we download and host the video, plays in app.

New system: Meta sends only the original Instagram URL. No downloadable file. No MP4. The reel lives on Instagram's servers and that's it.

We rebuilt the entire architecture in 48 hours.

New system: store the media ID, caption and thumbnail. When user taps to view, deep link directly to Instagram.

Honest take: it's actually better. We were hosting video files which meant storage costs, bandwidth costs, Firebase costs. Now it's just metadata and thumbnails. The bill is lower. The architecture is cleaner. Vestron was never meant to be a video player anyway. Instagram plays the reel. Vestron organises your life.

What I got wrong

I spent 6 months building before talking to a single user.

My first real user interview happened 2 weeks ago. One person. That's it.

That one interview told me the AI categorisation was defaulting to wrong categories when content was cross category. A bug I didn't know existed. Took 30 minutes to fix.

I should have talked to 10 users in month 1. I didn't. That's the honest gap.

What I got right

Event driven architecture from day one. All computation triggers from user saves, not scheduled jobs. 94.2% throttling efficiency on Firebase functions. Cost stays flat as usage grows.

The title system only evaluates once on Day 7 at the streak milestone. Cached for 20 hours. The system literally doesn't run unless it needs to.

What's next

Getting from 84 to 1,000 MAU.

First: the title card share feature. A Strava style transparent PNG sticker users post to Instagram Stories after their Day 7 reveal. That's the distribution engine.

Then intent capture, one tap chip on every save asking what's this for. Then Actualized %, did you actually do the things you saved. Then Brain Map, your saves visualised as an interest graph.

In that order. Not all at once.

The one thing I'd tell myself from 6 months ago

Talk to users before you build features. Even 5 conversations would have saved me months of going in the wrong direction.

Happy to answer anything. What are people building here that's in a similar space?