Mobile AI Assistant Offload Router

Arrival & Routing

Confidence threshold τ 0.65

Network RTT 120 ms

Query arrival rate 1.2 / s

Network dead-zone probability 10%

Live Readouts

On-device 50%Cloud 50%

Avg latency

0 ms

Battery drained

0.000 mAh

Avg answer accuracy

— %

Queries handled

Waiting for first query…

How it works

Every query a mobile AI assistant hears gets an on-device confidence score c ∈ [0,1] from a small local NLU model, then a router decides where to actually resolve it:

if network down (prob p_off):     resolve on-device (forced)
elif c ≥ τ (threshold):           resolve on-device
else:                              offload to cloud

latency_device = L0_dev + jitter
latency_cloud  = RTT + L0_cloud + jitter

accuracy_device(c) = 0.5 + 0.45·c
accuracy_cloud      = 0.97  (bigger model, always confident)

battery_device = E_npu           (fixed NPU energy / query)
battery_cloud  = E_radio_wake + k·RTT   (radio stays active for the round trip)

τ (confidence threshold) — raise it and more borderline queries get pushed to the cloud: latency and battery both climb, but so does accuracy on the queries that used to be answered locally with low confidence.
RTT slider — models WiFi vs. weak cellular; a high RTT makes every cloud round-trip both slower and more battery-hungry, since the radio stays awake for the whole trip.
Dead-zone probability — forces the router to answer locally even below τ, trading accuracy for availability, exactly like a real assistant losing signal.
The 3D view shows each query as a glowing instanced sphere: a short blue loop near the phone means it was answered on-device, a red arc out to the cloud node and back means it was offloaded.

Real-world relevance: this cascade — cheap on-device model first, expensive cloud model only when needed — is exactly how Siri, Google Assistant and Alexa keep common commands fast and private while still handling anything genuinely hard.

Resolved on-device Offloaded to cloud

⚠ Couldn't load the 3D engineThree.js failed to load from the CDN. Check your connection and reload.