All projects

2025 · Mobile, distributed backend and infrastructure

Yoma

A photograph flatters. Sixty seconds of video walking through a flat does not. Yoma makes video the listing, then does the tedious part of house hunting on the renter's behalf.

Visit the live site ↗yoma.dfnstudio.dev
Yoma
5
backend services, event driven
60s
of vertical video per listing
7
cities covered
3
databases, each for one job

The problem

Renting in Douala or Yaoundé means WhatsApp photos taken at flattering angles, agents who cost you an afternoon, and a trip across town to discover a place is nothing like its pictures. The wasted journey is the actual cost of the current system.

It is also asymmetric work: the renter has to contact every landlord one at a time, and most never reply.

What it does

  • Listings are up to sixty seconds of vertical video, not a gallery. You see the corridor, the water pressure, the street outside.
  • Search on a Google Maps view with filters that apply live, because in a city the question is always where before it is what.
  • An AI assistant reads the search and messages the five best matching landlords on the renter's behalf, so one search starts five conversations.
  • Direct secure chat between renter and landlord once contact is made.
  • Every landlord is verified by an admin before a listing can go up.
  • Douala, Yaoundé, Bafoussam, Dschang, Kribi, Limbe and Ngaoundéré.
  • An admin panel carrying the verification queue, listings, moderation and users.

How it is built

  • Five NestJS services: gateway, core, media, notifications and chat.
  • Completely event driven over BullMQ queues. Video transcoding, notification fan-out and AI matching all happen off the request path, because none of them should make a renter wait and all of them can fail and be retried.
  • Postgres for listings, users and everything that has to be consistent. Redis for caching and as the queue backbone.
  • MongoDB behind the chat service and nowhere else. Chat is the one part of this system with a fundamentally different shape: enormous write volume, append mostly, read by conversation. Giving it its own store keeps that load away from the database the listings depend on.
  • Media is its own service for the same reason: transcoding is CPU bound and bursty, and it must never be able to slow down a search.

Infrastructure

  • All five services containerised with Docker on a dedicated VPS, provisioned and maintained solo.
  • Splitting the backend was a decision about failure, not fashion: the parts that can be slow are the parts that are allowed to be slow, and they cannot take the rest down with them.

Stack

  • React Native
  • NestJS
  • BullMQ
  • Postgres
  • MongoDB
  • Redis
  • Docker
  • Google Maps

Other projects