One model remembers. Six forget on purpose.
The WhatsApp assistant behind Sukoon Unlimited, an agetech startup for elderly people seeking guidance, companionship and community. A stateful main model holds every conversation and delegates each task to a short-lived sub-model that does one job and disappears.
A kernel with six specialists
Seven hops, one reply
User writes on WhatsApp
The webhook hands the message to Flask.
MainGPT reads the context
The whole conversation is loaded from MongoDB.
It picks a tool
A function call names the sub-model it needs.
Controller spins it up
Fresh, stateless, scoped to this one query.
Sub-model fetches and answers
Calls its API, returns a structured result, and is gone.
MainGPT writes the reply
Results are merged into one answer in the user's thread.
Sent back on WhatsApp
History is saved; Slack hears about anything odd.
Why only one model keeps state
MainGPT
Owns the conversation. One place reads and writes history, so context never forks across models. Runs on gpt-4o.
Sub-models
Get a small prompt and a schema, call one API, return. Short prompts keep them cheap; adding a skill means adding one module.