The technology behind empathetic AI

Multimodal emotion recognition

AI USM analyses three signal channels at once: facial dynamics from the camera, prosody and timbre from the microphone, and the semantics of what a person actually writes or says. Each channel produces its own probability distribution over emotional states; a fusion layer weighs them against one another, so a smile that contradicts a tense voice does not silently win. On our internal validation sets the fused model reaches high accuracy, well above what any single channel achieves alone.

The pipeline is built for conversation speed rather than batch benchmarks. Inference runs under 200 ms end to end, which is what makes an avatar able to change tone mid-sentence instead of reacting a turn too late.

• Computer vision: micro-expression and gaze tracking with per-frame confidence scores
• Voice analysis: pitch, tempo, energy and pause structure, independent of language
• Text understanding: intent, sentiment and hedging cues in five languages
• Fusion layer: resolves conflicting signals instead of averaging them

AriGraph long-term memory

Most assistants forget the person between sessions. AriGraph, our proprietary memory layer, stores what matters as a graph of entities, events and relations rather than as a flat transcript, so the assistant can recall that a user's father had surgery in March without re-reading months of chat. Emotional context is stored alongside the facts: how a topic felt last time changes how it is raised this time.

Memory is scoped and revocable. A user can inspect what has been remembered and delete it, and family context is only shared between accounts that have explicitly linked.

Empathetic avatars with LipSync

Specialist avatars — medical consultant, fitness coach, nutrition and study companions — render in WebGL with speech-synchronised facial animation. The avatar's expression is driven by the same emotion state the model has inferred, so the visible reaction and the spoken answer agree. Where WebGL is unavailable the interface degrades gracefully to a voice-only mode.

AI USM assistants are decision support, not a medical device. They do not diagnose, and every health-related surface carries that disclaimer.

Privacy, residency and compliance

Emotion data is among the most sensitive categories a product can touch, so consent is explicit and per-modality: a user can allow voice analysis and refuse camera access. Raw video is processed for inference and is not retained by default; what persists is the derived, anonymised state.

The group operates through two entities: USM Tech Inc., a Delaware (USA) corporation, and AI USM LTD, registered in ADGM Abu Dhabi under number 28022 and endorsed by Hub71. The platform is built against GDPR and HIPAA requirements and the DOH Abu Dhabi framework for health-adjacent applications.

ULC Token infrastructure

ULC is the settlement token of the ecosystem, issued on Solana (mint FEr1ktFYTEmbFJocAqPQrhBotv1r5gCoXNgoEokNbqxD). It is used for assistant access, partner billing and ecosystem incentives; market data is read from GeckoTerminal and refreshed on a fixed schedule rather than on every page load.