SYMAR · Artificial intelligenceSince 2023
SYMAR: synthetic personas no longer depend on one language model
SYMAR ran its synthetic personas on one hard-coded language model. We rebuilt the architecture for four providers and let users pick the model for each task.
Project facts
Every engagement rests on a few checkable facts.
- Our role
- Architecture rework
- Industry
- Artificial intelligence
- Technology
- Multi-model LLM architecture (OpenAI, Google, Anthropic, Mistral) · Prompt chains · Multimodal agents for text, image and video
- Services
- AI agents & automation
Synthetic research had outgrown one model
SYMAR, formerly Opinio.ai, builds synthetic personas for market research, product testing and consumer-behavior analysis. A persona answers the way a defined customer segment would, and it works with text, images and video, not only with a questionnaire.
The platform ran every persona on one hard-coded language model. A short screening question and a long analysis of an uploaded video put very different demands on a model, and one model could not serve both well. Users had no way to choose, and response quality varied with the task.
What a researcher needs to decide
- Which model handles a task, weighed by quality, speed and cost
- How to switch providers without leaving the workflow
- How much context a persona gets, so its answers hold up
What the code did not allow
- The model connection was written into the architecture, not configured
- Providers expose different interfaces, and the code knew one
- Every change in the model market meant a change in the product
Model access moved out of the research workflow
We reworked the architecture so that models sit behind one common layer. Through it the platform connects to OpenAI, Google, Anthropic and Mistral. Adding a provider is an integration, not a redesign.
On that layer we built prompt chains: a larger request is broken into a sequence of instructions, and later steps use the context that earlier steps prepared. Specialized agents process text, image and video inside the context of the defined persona. The persona’s answer is assembled from those steps, not produced in one shot.
Last came the interface. A user picks the model for a task and switches it later, inside the same workflow.
One layer, four providers
The platform speaks to four providers through one integration layer. Prompt chains and the multimodal agents run on top of it, independent of which model is selected.
The user decides, not the codebase
Model choice is a setting in the interface, made per task. The user weighs quality, speed and cost. The product does not decide for them.
More choice, and no promise that one model fits every task
SYMAR now runs a multi-model architecture, prompt-chain workflows and a model-selection interface. What a persona answers depends on the sources and the steps behind it; which model runs those steps is the user’s decision. The collaboration continues.
Similar work, a different starting point
A description in your own words is enough to start.
Describe what is holding you back. Within two business days we get back to you with a first practical step.