Voice interaction
A custom wake word starts the loop. Speech is transcribed, answered by the configured AI provider, and spoken back
through a streaming text-to-speech pipeline — the first sentence is audible while the rest is still generating.
Sub-second streaming speech
Modular AI providers
Google Gemini by default, with NVIDIA NIM, Groq, OpenRouter and Cloudflare Workers AI available through one provider
registry. Adding a provider is a single registry entry.
5 providers · 101 tools
On-device vision
Face detection, multi-object tracking and 512-dimensional face embeddings run locally as ONNX models — recognition
does not require a network round-trip.
YuNet · ByteTrack · ArcFace
Object detection
YOLO-class detection identifies objects in the camera stream, with confidence thresholds and fine-grained labels
from a second on-device model.
Real-time · on-device
Memory
Conversations, entities and facts persist in a vector store, so Adalyn can recall context across sessions and
recognize people it has met before.
Vector · conversational
Physical control
Drive, climb, collect, gimbal and audio output on real REV hardware — behind a kill-switch safety system designed
to fail safe.
REV Control Hub
Companion experience
Pairing, enrollment, gallery review and settings from a phone or browser — the same design language as every other
Adalyn surface.
Flutter · mobile + web