synstate labs

Qualia

On-device context model for AI agents.
AI agents need to adapt to the person, the work, and the moment of the task. The model uses user context processed on device to produce compact context output for more relevant agent behavior.

Personalization

Personalization is the core problem of the next generation of AI agents.

An agent should understand who is working, what the person is trying to complete, what context matters now, and how the task is unfolding during the interaction.

The goal is practical personalization: fewer repeated explanations, fewer correction loops, and more completed work from the first useful response.

Context types

The model works with multiple types of user context. Behavioral signals are one part of that context.

Personal context
Preferences, recurring patterns, format expectations, and user-specific ways of working.
Work context
Current task, relevant project material, role context, internal terminology, and workflow state.
Behavioral signals
Signals from the interaction itself: hesitation, retries, corrections, context switching, and task continuation.
Interaction history
How the user has worked with the agent across previous attempts, corrections, and accepted outputs.
App and web activity
The surrounding work activity that helps identify what the user is doing now.
Task progress
Whether the task is moving forward, looping, waiting for clarification, or changing direction.

Model

The model turns user context into compact output for the agent run.

The model runs on device. Raw work material can be processed locally and reduced to a smaller task-specific output.

Input
Personal context, work context, behavioral signals, interaction history, app and web activity, task progress.
Model
On-device context model for the current user, task, and work situation.
Output
Compact context output for the agent run.
01
User context
02
Qualia on device
03
Compact context output
04
Personalized agent behavior

Demonstration

Corporate AI agent uses Qualia

Generic
stale facts2
Personalized
selected context5
Context
[1] Org

Org chart: CFO role changed

synced 14h ago · HRIS

[2] Org

Interim CFO: Dana Osei

synced 14h ago

[3] Decision

Cloud budget $1.2M → $1.45M

Tue leadership sync · meeting notes

[4] Terminology

Internal name: "Project Atlas"

214 mentions · 6 months

[5] Preference

Dana prefers bullets, <120 words

9 threads · conf 0.91

Where it fits

Qualia is most relevant wherever agents work repeatedly with the same user across real tasks.

AI agents
Personalize agent behavior with user context and task progress.
AI copilots
Adapt assistance inside ongoing work sessions.
Developer tools
Use project context, interaction history, and task progress in coding workflows.
Enterprise workflows
Personalize AI actions inside longer business processes.
AI product evaluation
Measure completion, corrections, retries, prompt length, token use, and repeated use.
Agent-driven products
Add user-context modeling where repeated use and task completion matter.

Evidence

Evaluation compares prompt-only agent runs with runs that include context output from the model.

Current measurements focus on completed work, repeated corrections, unnecessary clarification, prompt length, token use, and first-response acceptance.

−34%
Time to task completion
−41%
Retry and clarification rate
−47%
User-correction rate
−38%
Prompt length
62→84%
First-response acceptance
Internal evaluation. Research preview under customer-development validation.

Current stage

The current work is focused on model development, training, personalization, and evaluation of user-context signals in AI agent workflows.

We are testing how different forms of context change task completion, correction loops, prompt length, and first-response acceptance.

Model work

  • On-device context modeling
  • Personalization from user context
  • Behavioral and work-related signals
  • Compact context output

Validation

  • Corporate agent workflows
  • AI copilots and developer tools
  • Task-completion metrics
  • Selected beta discussions
Research preview — selected conversations with teams building AI agents and copilots.
Contact us