One day when I was using telebirr I thought when someone's new to the app, how do they handle it if they don't know how to use a feature? It only has a FAQ, and that's not exactly built for actually asking a real question in your own words.So I built a small RAG assistant that answers questions about telebirr sending money, cash-out at agents, failed transactions grounded strictly in what I actually wrote for it. I know it needs some improvement but for now I was thinking to scale to:

1.More data right now it's 5 hand-written Q&A pairs. Scaling means real data: telebirr docs, app reviews, maybe questions people actually ask.
2.A real vector database swapping my manual similarity search for ChromaDB, so it stays fast even with thousands of entries instead of just 5.
3.Amharic support the actual differentiator this needs, using a multilingual embedding model instead of English-only.
4.Better retrieval and Real evaluation

Let me hear your idea on the Improvement 🤝

I built a match maker for Ethiopian Telegram creators. Nobody asked. I did it anyway. It was a very fun build.

It reads public creator channels, builds a dating profile from what you post, and matches people by real chemistry, shared passions, matching vibes. 57 couples so far.

See why each couple matches, play matchmaker yourself, vote for your ships, and download match cards. The algorithm works in a way it maches using vibe, interest, content, tier, activity, and locality.

https://match.kidus-yohannes.engineer

(Matched and want out? DM me.)