The social intelligence platform for AI

The Citrus platform adds social intelligence to AI systems through a model-driven interaction stack. It can augment frontier models or operate as a standalone system. Specialized perception, user-modeling, and steering components interpret people over time and adapt AI behavior in real time.

Follow the interaction

One continuous interaction stack

Interpret What is happening with this user right now? Personalize Who is this user, and what matters to them? Steer How should the AI behave next?
Real-time perception

Interpretation Layer

Models the interaction as it unfolds across language, audio, vision, engagement, and software-event context. This layer captures not only what a user says but also how they say it and which moments matter.

SignalsMultimodal behavior and interaction context.
Temporal ContextTracks how interaction state changes across moments and turns.
StateProduces structured context that downstream systems can act on.
Auxiliary output: Objective interaction metrics & structured reports
Persistent user modeling

Rapport Layer

Builds a persistent representation of the person behind the interaction, separating relatively stable preferences and baselines from transient states, so that context can evolve with the user over time.

PreferencesLearns how users prefer to communicate and respond.
BaselinesTracks individual norms, so that behavior is interpreted relative to the person.
ContinuityMaintains longitudinal context across interactions instead of starting from zero.
Auxiliary output: Social-intelligence feedback
Adaptive behavior

Steering Layer

Uses live interaction states and persistent user context to adapt downstream AI behavior toward business objectives. The layer can modulate timing, communication style, interaction strategy, and task execution.

StrategySelects how the AI should approach the interaction based on the user and objective.
ExecutionAdapts language, timing, behavior, and task flow in real time.
ConstraintsCan incorporate product, safety, and user-level boundaries into the steering objective.

Defensibility

Interaction technology

Citrus develops proprietary models for human-AI interaction, alongside novel evaluation frameworks, for measuring and refining the models' behaviors. Current AI lacks objective metrics and labels for social intelligence, making the ability to measure social intelligence, and thus refine models accordingly, a technical advantage.

The platform in practice

One interaction, carried through the stack.

An illustrative diagnostic conversation shows how the layers shape the interaction as it unfolds.

User says
“I know I should schedule it. I just don't want to think about what the result could mean.”
Live signals Longer pause Speech slows Gaze shifts away
  1. Interpretation

    The interaction shifts from practical hesitation toward avoidance as possible outcomes enter the conversation.

  2. Rapport

    Relative to this user’s baseline, direct explanations build trust, urgency causes disengagement, and the exchange is moving away from readiness.

  3. Steering

    Match the user’s pace, lower pressure, and guide the exchange toward one manageable decision.

Interaction behavior
Brief pause Slower delivery Lower conversational pressure
Adapted AI response

“That uncertainty can make even scheduling feel heavy. We can take this one step at a time—would it help to start with what the appointment actually involves?”