aiQu Labs

ai · INDUSTRIAL & PHYSICAL AI

AI that runs where the work happens.

Perception, prediction and control for machines, lines and fleets — built end to end, from sensor to cloud.

Talk to an EngineerSee the custom AI pipeline

THE GAP WE ENGINEER FOR

The model is rarely the hard part.

Most industrial AI stalls between the notebook and the plant floor. The model works on a sample set, then meets dust, vibration, bad lighting, a 12-watt power budget and an operator who has 90 seconds. We engineer for that gap first: hardware, latency and failure behaviour are design inputs, not afterthoughts.

CAPABILITIES

Five capability areas across the ai vertical.

Most engagements combine two or three. All of them ship as working systems, not reports.

01

Physical AI & Edge Intelligence

Perception and decision-making on devices — smart cameras, robots, drones and machines. On-device inference on embedded SoCs and NPUs, sensor fusion, and real-time control loops.

  • Model quantisation and compilation for the target NPU
  • Camera, LiDAR, IMU, thermal and acoustic fusion
  • Closed-loop control with deterministic latency
  • Offline-first operation with store-and-forward sync
02

Industrial Vision

Defect detection, quality inspection, PPE and safety compliance, OCR on production lines, plus counting and tracking.

  • Defect taxonomy built with your quality team
  • Evidence frames stored per unit for traceability
  • OCR and code reading on moving product
  • Counting, tracking and zone analytics
03

Predictive & Prescriptive Maintenance

Vibration, thermal and acoustic analytics, anomaly detection, and remaining-useful-life models for rotating and reciprocating assets.

  • Retrofit sensor kits where instrumentation is thin
  • Per-asset baselines instead of fleet thresholds
  • Remaining-useful-life once failure history exists
  • Recommended action, not just an alarm
04

Industrial Copilots & Agentic AI

LLM agents grounded in SOPs, manuals and plant data: maintenance assistants, deviation triage, and shift-handover summaries.

  • Retrieval grounded in your document set, with citations
  • Agents that read historian and MES data
  • Shift handover and downtime summaries
  • Guardrails and an audit log on every answer
05

End-to-End Custom AI

From problem framing and data collection to custom models, hardware design, firmware integration, deployment and lifecycle support — one accountable team, and you own the IP.

  • Single team across hardware, firmware and models
  • Fixed-scope discovery before any build commitment
  • Source, weights and documentation handed over
  • Support and retraining sized to your ops team

GO DEEPER

Three pages, three levels of detail.

HOW WE DELIVER

Same six stages, every build.

Discovery is fixed-scope and ends in a go / no-go. Nothing is built before the baseline is measured.

  1. Step 01

    Discover

    Walk the floor, watch the process, and write down the decision the model has to make. Scope, constraints and a success metric before code.

    • Problem brief
    • Success metric

  2. Step 02

    Data

    Instrument what is missing, label what exists, and build the capture rig. We design the dataset for the failure modes that matter.

    • Capture rig
    • Labelled set

  3. Step 03

    Model

    Train against a measured baseline. Architecture chosen for the latency and power budget of the target device, not for a leaderboard.

    • Benchmarks
    • Model card

  4. Step 04

    Edge & Hardware

    Quantise, compile and integrate: board selection, sensor interfaces, firmware, thermals and enclosure. The model meets real silicon here.

    • Board bring-up
    • Firmware

  5. Step 05

    Deploy

    Staged rollout on one line or one fleet, shadow-mode first. Operators see what the system sees, and can override it.

    • Pilot line
    • Operator UI

  6. Step 06

    MLOps & Support

    Monitoring for drift, scheduled retraining, OTA updates and an on-call path. Handover docs so your team can run it alone.

    • Drift monitors
    • OTA pipeline

Scroll sideways to see all six stages.

WHERE IT APPLIES

Sectors already running these systems.

FAQ

Questions engineers ask first.

Do we need to replace our existing cameras or PLCs?

Usually not. We start by auditing what is installed and what it can stream. Where existing hardware cannot meet the resolution, frame rate or trigger timing a use case needs, we say so and scope the specific replacement — not a rip-and-replace.

How much data do we need before starting?

Less than most teams expect for a pilot, more than most expect for production. A first inspection model often needs a few hundred labelled examples per defect class. Predictive maintenance needs failure history, which is why we usually begin with anomaly detection and add remaining-useful-life later.

Will this run without internet connectivity?

Yes. Edge deployments run inference locally and buffer results when the link drops. Cloud is used for retraining, fleet dashboards and updates, not for the decision in the loop.

Who owns the models and the data?

You do. Weights, source, training configuration and documentation are handed over. We do not resell your data or train shared models on it.

NEXT STEP

Bring us a line, an asset or a fleet.

We will tell you what is feasible on your current hardware, and what is not.

Talk to an EngineerExplore Qu