Skip to content

AI for Physical AI

Judgement at line speed, without the misses.

We engineer AI for manufacturing and warehouse operators: line-side inspection, pick-and-place perception, warehouse robotics, and OEE monitoring.

Built on your own line data, your SCADA stack, and the functional-safety posture your plant already runs.

Discuss your line problemSee how the system works

The operational reality

Your line runs on takt. Takt runs on inspection. Inspection runs on humans.

At speed, humans miss. That defect ships. The customer finds it.

The camera doesn't blink. The customer doesn't forgive.

Felt by VP Manufacturing, Heads of Quality and Automation, Plant Managers, and the Industrial AI leaders carrying the OEE number.

  1. 01

    Line-side inspection

    Operators and cameras check each unit against spec at critical stations; defects are either reworked or scrapped.

    Friction · Human inspection drifts across a shift; bolt-on cameras raise alarms without context.

  2. 02

    Pick-and-place and sortation

    Robots pick parts from bins and totes and place them for the next operation; sortation routes units to the right lane.

    Friction · Novel SKUs break the vision model that worked last quarter; recovery takes an engineer on the floor.

  3. 03

    Warehouse perception

    Depalletising, case-sorting, and AMRs move inventory between zones against a plan that updates by the hour.

    Friction · Mixed pallets, damaged cartons, and label occlusion collapse into human-handled exceptions.

  4. 04

    SCADA and OEE monitoring

    PLC and SCADA signals land continuously; OEE is reported shift by shift against targets.

    Friction · Downtime causes are logged after the fact; by the time the pattern is clear the shift is already gone.

  5. 05

    Audit and traceability

    In regulated manufacturing, every reject, every rework, and every operator action is recorded for inspection.

    Friction · The audit trail is spread across the MES, a spreadsheet, and a camera NVR nobody wants to search.

Swipe the workflow

Where the work breaks

The friction is in the places nobody budgets for.

  • 01

    Human inspection drifts across the second half of the shift.

    How we fix it · A system we build catches this before it costs you.

    Impact · Defect escapes rise after lunch; the recall risk is in units that shipped on Friday afternoon.

  • 02

    Bolt-on cameras alarm without the context the operator needs to act.

    How we fix it · A system we build catches this before it costs you.

    Impact · Alarms are muted after the second false reject; the station runs blind.

  • 03

    Novel SKUs break vision models trained on last quarter line.

    How we fix it · A system we build catches this before it costs you.

    Impact · First-pass yield collapses on a new product; recovery pulls an automation engineer to the floor.

  • 04

    AMRs stall on mixed pallets, torn labels, and shrink-wrap glare.

    How we fix it · A system we build catches this before it costs you.

    Impact · A perception exception becomes a human-handled move; the hourly plan slips.

  • 05

    Downtime causes are logged after the fact.

    How we fix it · A system we build catches this before it costs you.

    Impact · OEE improvement runs a week behind reality; the same cause recurs across shifts before planning sees it.

What we would build

Systems, not features. Each with its own job.

  • 01

    Line-side inspection with calibrated confidence

    Cameras alarm without context; operators mute them; the station runs blind.

    Who benefits · Line operators, quality engineers, and plant managers.

    Inputs
    Line cameras, part geometry and tolerance from engineering, SKU change-over state from MES, and a reviewed library of past defects.
    What the system does
    A per-station vision model trained on your own good and defect imagery scores each unit against spec, with bounding boxes and per-detection confidence; low-confidence calls are routed to a QA review lane rather than auto-rejected.
    Output
    A pass or fail decision per unit, with the defect type, the measurement, and the model confidence; every rejection carries the imagery for audit.
    Benefit
    Defect-escape rate drops; false-reject rate drops with it; the station stops muting alarms.
    Human stays here
    Operators approve, override, or escalate every borderline call. A QA lane reviews a sampled share of pass decisions on an ongoing basis.
  • 02

    Pick-and-place perception for novel SKUs

    A new SKU breaks the vision model the robot has been running against for months.

    Who benefits · Automation engineers, line engineers, and the SKU introduction team.

    Inputs
    CAD or product imagery for the new SKU, a short on-floor capture pass, the robot kinematics, and the bin geometry.
    What the system does
    A perception stack that learns the new SKU from a short capture pass rather than a from-scratch training cycle, with uncertainty flagged on each pick for a slow-mode fallback.
    Output
    A per-SKU pose estimate and grasp plan, with slow-mode and human-pick fallbacks for low-confidence frames.
    Benefit
    SKU introduction stops pulling an automation engineer to the floor for days; first-pass yield recovers inside a shift.
    Human stays here
    A line engineer approves the first batch after introduction; slow-mode and human-pick lanes are first-class, not failure states.
  • 03

    Warehouse perception for depalletising and sortation

    Mixed pallets, torn labels, and shrink-wrap glare collapse into human exceptions.

    Who benefits · Warehouse operations, automation leads, and the AMR fleet owners.

    Inputs
    Pallet and case imagery, label OCR, WMS pick plan, AMR fleet state, and a library of past exceptions.
    What the system does
    A perception stack built for warehouse mess (occlusion, glare, damage) with explicit exception handling that routes to a human pick lane rather than failing silently.
    Output
    A pick, place, or hand-off decision per unit; AMR routes that respect current exception state; a per-exception reason for the fleet log.
    Benefit
    Human-handled exceptions drop; AMR throughput rises; the hourly plan holds across the shift.
    Human stays here
    Human pick lanes are a routing target, not a failure. Fleet dispatch retains an override on any route.
  • 04

    SCADA-aware downtime causation

    Downtime causes are logged after the shift, when the pattern is already gone.

    Who benefits · Plant Managers, Industrial AI leads, and shift supervisors.

    Inputs
    PLC and SCADA signals, OEE event streams, maintenance history, and operator notes from the shift.
    What the system does
    A causation model clusters downtime events across lines and shifts, surfacing recurring root causes in near real time rather than at the weekly review.
    Output
    A live downtime causation view with the top recurring causes, the affected stations, and a drill-down into representative events.
    Benefit
    OEE improvement moves from weekly to live; the same cause stops recurring across shifts.
    Human stays here
    Shift supervisors confirm or reclassify every surfaced cause; model classifications never overwrite operator notes.
  • 05

    21 CFR Part 11-aware audit trail

    The audit trail is spread across MES, spreadsheets, and a camera NVR nobody wants to search.

    Who benefits · Quality leaders, compliance leaders, and plant managers in regulated manufacturing.

    Inputs
    Inspection decisions, operator actions, model versions, SKU and lot data, and camera imagery with retention policies.
    What the system does
    A unified, signed, immutable record of every inspection decision, every operator override, and the model version in use at the time, retrievable by lot, SKU, or station.
    Output
    A single-pane audit trail with signed records, model-version provenance, and imagery references per decision.
    Benefit
    Audits that took weeks compress into hours; recall traceability lands within the regulatory window.
    Human stays here
    Signatures are performed by authorised operators under their own identity; the system records, it does not sign on anyone behalf.

Show the system

What it could look like in your hands.

A design concept, not a shipped product. Interfaces like this get drawn out before any code lands, so the system reads back the way the people using it work.

inspection / line review

Focused inspection

Station 04B · SKU-7310Fail

Lot L-20261009-0417 · model confidence 94%

Surface scratch, 1.2mm · 97%Surface deviation, 0.4mm · 86%
  • Surface scratch, 1.2mm97%
  • Surface deviation, 0.4mm86%
Synthetic data. This is a design concept of how a shift feed, a focused inspection with detections, and an actions pane could sit together for a line operator.

End-to-end workflow

From the trigger to the outcome.

A representative path from a unit arriving at a station to a signed audit record. Automated stages run at line speed; human stages remain explicit and recorded.

  1. 01

    Unit arrives at station

    The MES signals the SKU and the lot; the station cameras capture the unit; SCADA state is tagged to the frame.

    Integration
  2. 02

    Perception runs

    The per-station vision model scores the unit against spec and returns bounding boxes with per-detection confidence and an overall pass or fail call.

    Automated
  3. 03

    Confidence gate

    Low-confidence calls are routed to a QA review lane rather than auto-rejected; borderline frames are stored for retraining.

    Guardrail
  4. 04

    Operator confirms

    The operator approves, overrides, or escalates each borderline call; the UI shows the detection, the measurement, and the confidence.

    Human in loop
  5. 05

    Decision written

    The decision, the imagery reference, the model version, and the operator identity are written to the MES and the audit store.

    Integration
  6. 06

    SCADA feedback

    Pass-rate and defect-type trends feed back into the OEE and downtime causation views in near real time.

    Automated
  7. 07

    Audit trail available

    A lot or SKU query returns every decision, every override, and the model version in use at the time, as a signed, retrievable record.

    Integration

How exceptions are handled

  • If confidence is below threshold, the unit routes to the QA review lane rather than being rejected or passed silently.
  • If a new SKU is introduced without an approved perception model, the station runs in slow mode until a line engineer approves.
  • If a camera stream drops, the station pauses rather than inspecting on partial data.
  • If an operator overrides a decision, the override is recorded with reason codes and surfaces in the shift review.

How it is engineered

Built on your stack, not ours.

Design choices that let perception run at line speed, inside your plant network, under your functional-safety posture.

01

Edge inference

Perception runs on an edge compute node at the line, with bounded latency and no reliance on a cloud round-trip for pass or fail decisions.

02

SCADA and MES integration

PLC, SCADA, and MES signals are read through the plant OT boundary via your approved interfaces; writes go through MES, not direct to the control layer.

03

Perception model lifecycle

Per-station, per-SKU models with explicit retraining windows, change-management sign-off, and an evaluation harness before any model promotes to the floor.

04

Identity and access

SSO through your existing identity provider; operator identity is first-class on every decision; service accounts never sign.

05

Data boundary

Camera imagery and SCADA data stay on your network. Where regulated, retention windows and signature records meet 21 CFR Part 11 and ISO 13485 expectations.

06

Functional safety

The system is designed as a quality and routing layer, not a safety-rated controller. Safety-rated interlocks remain with the PLC and the safety PLC under ISO 13849 and IEC 61508.

What the business will measure

Numbers the operator already watches.

We agree on metrics and baselines before building. These are the numbers a plant usually watches when deciding whether a perception system is working.

  • 01

    Defect-escape rate

    Baseline today

    Share of defective units reaching the customer, by SKU and station.

    How it moves

    Calibrated vision with a QA review lane catches the defects human inspection drifts past.

  • 02

    First-pass yield

    Baseline today

    Share of units passing inspection on first attempt, by station and shift.

    How it moves

    Early flagging and SKU-aware perception reduce rework and scrap across the line.

  • 03

    False-reject rate

    Baseline today

    Share of units rejected that are actually within spec, by station.

    How it moves

    Per-detection confidence and a QA review lane cut false rejects and the operator mute response.

  • 04

    Inspection throughput

    Baseline today

    Units inspected per hour against takt time, by station.

    How it moves

    Edge inference holds inspection latency inside takt, so throughput does not trade against quality.

  • 05

    OEE lift

    Baseline today

    Overall Equipment Effectiveness per line, decomposed into availability, performance, and quality.

    How it moves

    Live downtime causation plus SKU-aware perception reduce availability loss and quality loss in parallel.

Human in the loop

Oversight, safety, and the regulatory posture

Where the humans stay in control.

A line answers to quality, to compliance, and to the operators who work it. The system is designed around that posture, not against it.

  • 01

    Quality management

    Compatible with ISO 9001 at the quality-management level and ISO 13485 for medical-device manufacturing; signatures, model versions, and audit trails are first-class.

  • 02

    21 CFR Part 11 posture

    In regulated manufacturing, electronic records are signed, immutable, and retrievable by lot and SKU; operator identity is enforced on every signature.

  • 03

    Functional safety

    The system is a quality and routing layer, not a safety-rated controller. ISO 13849 and IEC 61508 interlocks remain with the PLC and safety PLC.

  • 04

    Workplace safety

    OSHA and EU Machinery Directive posture is respected: no action is taken that overrides a lockout or a safety interlock, and emergency stops always win.

  • 05

    Attribution and audit

    Every inspection, every override, every model promotion logs the model version, the inputs, and the operator or engineer who acted.

  • 06

    Low-confidence routing

    Borderline decisions route to a QA review lane rather than being pushed through; the system refuses to pass or reject when confidence is below threshold.

Proof

Work we have shipped, and the shape of what we build.

Real engagements where we can speak to them, and representative use cases where we cannot.

  • Use case

    Calibrated line-side inspection

    A proposed design: per-station vision with per-detection confidence and a QA review lane, catching the defects human inspection drifts past without raising the false-reject rate.

  • Use case

    Fast SKU introduction for pick-and-place

    A proposed design: a perception stack that learns a new SKU from a short on-floor capture pass, with slow-mode and human-pick fallbacks for low-confidence frames.

  • Use case

    Live OEE causation

    A proposed design: a causation model that clusters downtime events across lines and shifts in near real time, so the same cause stops recurring before the weekly review.

How the build runs

Scoped tightly, shipped in phases.

Fixed price per phase, approved before the phase starts. You own the code, the models, the data, and the deployment.

  1. 01

    Discovery

    Two weeks. We walk the line, map your MES, SCADA, and quality posture, and agree on the metrics we will move and how we will baseline them.

  2. 02

    Prototype on real data

    Four to six weeks. One station, one SKU family, one proof of mechanism. Shadow-run against your line; we measure against the baseline.

  3. 03

    Production rollout

    Eight to twelve weeks per line. Edge compute, MES and SCADA integrations, SSO, audit trails, evaluation harness, and the operator workflow your quality team signed off on.

  4. 04

    Handover

    You get the full source, the runbooks, the evaluation suite, and a two-week walkthrough with your automation, quality, and compliance teams.

  5. 05

    Optional ongoing support

    A fixed-term support agreement for model-drift monitoring, SKU introductions, and plant-floor change management. Never a retainer dependency.

Common questions

Answers before the first call.

01Does this control the line?
No. The system is a quality and routing layer. Safety-rated interlocks remain with the PLC and safety PLC under ISO 13849 and IEC 61508; emergency stops always win.
02What about 21 CFR Part 11 and ISO 13485?
In regulated manufacturing, electronic records are signed, immutable, and retrievable by lot and SKU; model versions and operator identity are enforced on every signature.
03What happens when a new SKU is introduced?
The perception stack learns the new SKU from a short on-floor capture pass. The station runs in slow mode with a human-pick fallback until a line engineer approves the new model.
04Where does inference run?
On an edge compute node at the line, with bounded latency and no reliance on a cloud round-trip for pass or fail decisions. Training and evaluation can run centrally.
05How do you handle false rejects?
Per-detection confidence drives routing: borderline calls go to a QA review lane rather than being auto-rejected. Operators can override and the override is recorded with a reason code.
See the other details on the FAQ page

Next step

A thirty-minute conversation. Bring one station that loses you yield or inspection throughput today, and we will walk through what the first phase would look like, where the measurements come from, and what has to be true for your quality team to approve it.

Discuss your line problem

Helpful to have on the call

Helpful to have on the call: your MES and SCADA stack, one station with a yield or inspection pain point, and the person who owns the metric you would like to move.

Explore

A different problem in every sector, one engineering standard underneath.

Featured · AI for healthcare

Healthcare

Why is your clinician still charting at midnight?

Your clinicians are charting until midnight and your front desk is drowning in prior auths. We build AI that writes clinical notes automatically, triages patients to the right care faster, and finds any record with a simple question, so your people spend their hours on patients instead of paperwork.

See AI for healthcare
stallwart / healthcare

Live run

Healthcare

Running
as of nowhealthy
See every industry we build for

An AI-First Engineering Company

Get a price in one call.

One 30-minute call. You leave with an architecture, a fixed price, and a start date. No charge, and no slide deck.

Booking now for December 2026

Book a callSee how it works