Work
A multi-channel AI concierge, in production
In production: a live AI concierge answering customers on web, Telegram, Messenger, Instagram and WhatsApp, with cost-aware model routing and a staff dashboard that shows spend per channel.
- What it proves
- That we run AI in production and track what each channel costs.
- Method stage
- Built from
- Multi-provider model routing with cost-aware fallback
- Web, Telegram, Messenger, Instagram and WhatsApp channel connectors
- Conversation summarisation
- Staff dashboard: question clustering and per-channel cost estimation
- Sectors
The problem
Customers use the channel that is already open on their phone, and a concierge that lives only on a website misses the rest. Five channels also raise a cost question: what does one conversation cost, per channel, at volume, month after month?
What we built
Web chat, Telegram, Messenger, Instagram and WhatsApp connect to one reasoning and knowledge layer. A customer gets the same answer on any of them. Conversations are summarised so context carries forward.
Requests are routed across several model providers with cost-aware fallback. When a provider has a bad day or a channel spikes, traffic moves to another route, and the budget holds.
The staff dashboard has two jobs. It clusters the questions customers ask, so staff see what comes up most and where the assistant struggles. It estimates cost per channel, so “is this channel worth running” has a number behind it.
What it proves for a client
Routing survives a provider failure, cost is visible per channel from day one, and staff see where the assistant falls short. It is the only AI system we currently run in production.
Method stages
It is the nearest thing we have to Stage 06 · Run and evolve: a system that is operated, watched for cost and fed by what customers ask. See the KARPINITY Method for the earlier stages.