← Innovation Labs/Lab L4BetaEdge AI & sensing

Ambient Intelligence Fabric — Edge NPU · TinyML · Multimodal · Privacy-by-default

Round-tripping every frame to the cloud is dead. Ambient Intelligence Fabric runs vision, audio and signal models on edge NPUs with sub-10ms latency, sovereign by default — only the inference, never the raw stream, leaves the device. Privacy and physics, finally aligned.

Research thesis

Sovereignty, latency and unit economics all point to inference at the edge for ambient workloads.

Edge inference p95

9 ms

Bandwidth saved

− 96%

Cost per inference

− 89%

Become a design partner →See live experiments

Active experiments

What the lab is testing right now

Quantised VLMs on consumer NPUs

INT4 / INT8 vision-language models on Snapdragon, Apple Neural Engine and Hailo.

Federated learning loops

Models improve from edge feedback without raw data ever leaving the device.

Multimodal sensor fusion

Vision + audio + IMU on a single NPU pipeline for safety and ops use cases.

WebGPU as edge runtime

Browser-side inference for laptops, kiosks and POS — zero install.

Shippable artefacts

Everything the lab ships

Edge runtime
ONNX + WebGPU runtime with hot-swap models, signed deploys, observability.
Quantisation toolkit
INT4 / INT8 / mixed-precision pipelines with eval-preserving guardrails.
Federated trainer
Aggregated gradients only; differential privacy budget per round.
Privacy receipts
Per-inference cryptographic receipt: what ran, where, on what data class — never the data itself.
Reference deploys
Retail loss-prevention, factory safety, contact-centre ambient and fleet POS patterns.

Lab team

Edge AI Principal
Embedded / NPU Engineer
Quantisation Researcher
Privacy Engineer

Partners we collaborate with

NVIDIA JetsonHailoQualcomm AI HubApple Core MLONNX RuntimeWebGPU W3C

Example output · Manifest · edge.node.factory-7yaml

node: factory-7.bay-3
hardware:
  npu: hailo-8l
  cpu: arm-cortex-a78
  ram_gb: 8
models:
  - id: vision.safety.helmet_v6
    quant: int8
    p95_ms: 7
  - id: vision.safety.proximity_v3
    quant: int4
    p95_ms: 9
runtime:
  privacy: on_device_only
  receipts: signed
  uplink: aggregated_metrics_only
  bandwidth_budget_mb_day: 12
federated:
  enabled: true
  dp_epsilon: 1.2
  rounds_per_day: 4

Engagement timeline

Weeks 1–10 · first edge node live by week 3

1
Weeks 1–3
Edge node baseline
Hardware bring-up, signed deploy, telemetry, first quantised model live.
2
Weeks 3–6
Multimodal fusion
Vision + audio + signal fused; privacy receipts on every inference.
3
Weeks 6–10
Federated loop live
Models improve from edge feedback under DP budget; uplift measured.

Flagship pods

Productionised by these squads

Retail Loss-Prevention Pod

Factory Safety Pod

Contact Centre Ambient Pod

Fleet & POS Edge Pod

Selected publications

Receipts, not just thesis

Sub-10ms multimodal inference on consumer NPUs
MLSys Workshop·2025
Federated edge fleets without raw data: a 6-site field study
AXP Internal Whitepaper·2026

FAQs

What partners actually ask

What hardware do we need?

We support Hailo, Apple Neural Engine, Snapdragon, NVIDIA Jetson and WebGPU laptops. Choice is policy, not lock-in.

How is this private?

Raw streams never leave the device. Only signed inference receipts and aggregated metrics uplink — by default.

Federated learning safety?

Differential privacy budget per round, gradient clipping, server-side aggregation only — no raw data, ever.

What about model drift?

Edge eval harness runs nightly; drift triggers re-deploy or shadow mode automatically.

Design-partner programme · L4 Ambient Intelligence Fabric

Co-build Ambient Intelligence Fabric with us in Weeks 1–10.

We'll respond within one business day with a scoping note, a fixed-price outcome contract, and a named principal cleared for your domain. Design partners get first-look access, joint publication rights and roadmap influence.

• Outcome-priced — no T&M.
• Sovereign by default — your data, your region, your keys.
• Refund-backed if the contracted KPI isn't hit.
• Joint publication rights and conference slots.