The Business & Entrepreneurial Profile

With Maia's new backend, we've shifted from the user's entire conversation history as context (truncated to fit into the LLM's window) to building a more intentional business & entrepeneurial profile and providing that with a limited number of recent conversation turns to the LLM. This profile is the heart of the LLM's context, and has an number of advantages:

1) As conversation histories become long (we now have users with >1yr of history) the truncation can cut important business information provided in early turns.

2) Research shows LLMs can get lost in long multi-turn convesations. Reducing the conversation history helps reduce those episodes.

3) Even for users with shorter conversation history, our testing shows better results with this two-stage process: first, a dedicated profile builder (with scheduled and episodic rebuilds), then second, the main LLM using this highly efficient and strucutred profile to respond.

This approach also shows how a tool dedicated to helping MSMEs can perform better than a generalized tool. GPT 5's sytem prompt reveals a similar structure: it builds up a 'user bio' over time to contextualize future responses. But that bio is structured to give contextualized responses about health, coding, cooking, family travel, and everything else ChatGPT is used for. We have the luxury of building a business and entrepreneurial profile highly targeted on a single purpose: to help business owners sell more, spend less, and grow. Since this is the only thing we do, we can explicitly keep out all PII and all the stuff advertisers care about but a business mentor would not. This is also where the research into business training and MSME productivity interventions becomes highly relevant- we include key fields like what is the businesses' stage of formalizastion, is it banked, and what are the entrepreneur's key stressors, etc. And we can get the responding LLM to fill in key parts of the profile with targeted questions over time. There are also key fields for particular sectors, like coffee growers and restaurants.

The resulting profile, along with a more efficient record of past conversations, actions, and outcomes, greatly improves the tool's responses in our internal testing. We continue to refine the profiles and its integration into conversation history in a phased roll-out, and will share the structure in the coming weeks when that is complete.