Case study · consumer & community

MeshTribe: 8 product surfaces on one backend, and an AI layer above it

The digital home for motorcycle riders: clubs, rides, events, member content, a marketplace and notifications, with a mobile app and a web front end running on one backend. Live on web, iOS and Android.

Python · FastAPIFlutterPostgreSQLLiveAgent layer · acceptance testing
8
PRODUCT SURFACES
22
DOMAIN MODULES
3
PLATFORMS LIVE: WEB, IOS, ANDROID
1
BACKEND CONTRACT
The challenge

Every feature of a community platform is deterministic software working correctly. The real workload is judgement about human content: whether a profile is genuine, whether a report is serious, what deserves a member’s attention, what is worth writing up. That work arrives faster than a small team can review it.

The platform had to carry the community first, and then carry an AI layer that helps the people running it without ever acting on a member by itself.

What we built
  • 22 domain modules across identity and machines, community (clubs, gallery, stories, news), activity (events, rides, tours, training), services and commerce (products, mechanics, tow) and platform (notifications, scheduler, insights, admin).
  • One backend contract serving the web front end and the Flutter mobile application, so a feature ships once and appears everywhere.
  • Role-based access and privacy enforced centrally rather than per feature, a full lifecycle engine with recurring scheduling, rule-enforced club membership and polls, an admin review surface, and push delivery.
  • An AI layer designed onto the platform, in acceptance testing: profile integrity (is this a real rider, is the imagery authentic), content and report triage (classified against community policy, ranked by severity, the offending passage cited), community editorial (articles and a newsletter drafted from real activity, every claim traceable), and relevance (notifications and marketplace matches from a per-member interest profile).
  • The agents recommend and never enforce. Every verdict carries reasons, evidence and confidence; every run lands in an audit ledger; and irreversible actions, ban, suspend, delete, publish, send, stay with a person.
How the AI layer behaves

Reads, judges, drafts and recommends. People act.

The layer attaches at a defined seam and is labelled acceptance testing wherever it appears on this site, because it is not yet live.

Thirteen candidate capabilities were cut to four. Nine were consolidated, three were rejected because a rule expressed them better than a model would, and one was reserved permanently for people. Knowing what not to automate was most of the design.

Running a platform with content a team cannot keep up with?

Sixty minutes on where an agent layer would help, what it must never do alone, and how it attaches without destabilising what already runs.

A 60-minute architecture review · no charge · the notes are yours either way