Platform

Visual inspection AI that QA can release.

Teach from conforming units. Detect deviations nobody defined. Trace each one to its batch and component lot. Keep a record that stands up to an inspection.

Register
RUNNING
Simulated
inspection
PASS
PASS
REJECT
PASS
PASS
pass 0reject 0
on-prem inference · frames retained 30 d
Inspection recordbatch 24-0917
  • Awaiting units…
e-record · audit trail on · sha256 per frame
Investigation signal

No correlated pattern in this batch yet.

P-01 / Few-shot teaching

A few dozen good units is a training set.

Phamqyra ships with a vision foundation model pretrained on pharmaceutical containers and packaging — glass, elastomers, liquids, foils, print. Teaching a new product adapts that prior to a few dozen of your own conforming units, captured on your station. The result is a locked, versioned model of a conforming unit, ready for qualification.

  • Self-supervised: no defect labels needed
  • Handles natural variation in fill, meniscus, cake and print
  • One locked model per product, presentation and station
learn · 2R-clear · station 3simulated
01
02
03
04
05
06
07
08
09
10
11
12
Model 2R-clear-v7 ready
24 conforming units · self-supervised · sha256 3f9a…c21e
Submit for qualification

P-02 / Zero-shot deviation detection

If it deviates from conforming, it is flagged.

Every frame is scored against the learned representation. Deviations are localised to the pixel, scored and clustered. A deviation type nobody has defined still registers on its first occurrence — then recurrences are grouped so QA can review and name the class.

  • Pixel-level deviation maps with scores
  • Clustered deviation classes, reviewed and named by QA
  • Acceptance thresholds per product, under change control
detect · batch 24-0917 · side camsimulated
deviation mapscore 0.96
  • Visible particle
    class named by QA
    0.96
  • Glass crack
    class named by QA
    0.92
  • Unclassified cluster C-07
    first occurrence · needs review
    0.87
  • Low fill volume
    class named by QA
    0.85

P-03 / Batch investigation

Deviations, tied to the batch and lot that caused them.

Each result carries batch, sub-lot, station, shift and component lots. Phamqyra runs control charts per deviation class, surfaces shifts as they emerge, and drafts deviations into your QMS with the frames and statistics attached.

  • Correlation by batch, stopper, glass and label lot
  • Draft deviations to QMS, results to MES and LIMS
  • False-reject and escape trends per product
investigate · visible particle · batch 24-0917simulated
Where “visible particle” concentrates
lift vs. site baseline
  • Stopper lot ST-22912.8×
  • Filling needle 4 · after 14:10 changeover1.9×
  • Glass lot G-55301.2×
  • Batch 24-0915 · baseline1.0×
Draft deviation · DR-24-0412 · sent to QMS

Visible particle rate is 2.8× baseline for units stoppered from lot ST-2291. Recommend quarantine of remaining stoppers pending review.

Architecture

Under the hood.

A foundation model, a deviation engine and a model registry — packaged for people who run validated lines, not GPU clusters.

Pharma vision foundation model
Pretrained on containers, closures, liquids and packaging so few-shot adaptation converges quickly and robustly.
Deviation scoring engine
Patch-level representations compared against the learned conforming manifold; localisation without segmentation labels.
Review loop
QA dispositions feed new versions of thresholds and deviation classes — through change control, never silently.
Model registry
Every model versioned with its training units, metrics and validation evidence. Promote, retire and roll back with signatures.
Edge runtime
Inference on an appliance beside the machine, keeping pace with the line with no cloud round-trip.
System integration
OPC-UA and REST for machines, MES, LIMS and QMS. File-drop and historian connectors on request.
Validation evidence
Protocol templates, executed test evidence and a traceability matrix generated per model version.
Roles for the site
Operators run, QA dispositions and releases, engineers configure, auditors read. One workspace with segregation of duties.

04 / Compliance

Built for GMP, not retrofitted for it.

AI that QA can actually release: locked, versioned models, a validation package per version, and records that stand up to an inspection.

model lifecycle · 2R-clearv7 released · v8 in qualification
01Teachdraft model · v802QualifyIQ / OQ / PQ evidence03Releasee-signed · locked04Monitordrift · trending05Change controlreason for change0 untracked model changes
Audit trail by default
Every action attributable and time-stamped, with reason-for-change captured. Designed to support 21 CFR Part 11 and EU GMP Annex 11 workflows.
Electronic signatures
Model release, threshold changes and deviation dispositions are signed with role-based approval.
Validation package
URS, risk assessment, IQ/OQ/PQ protocol templates and a traceability matrix ship with every release.
Locked models
A released model does not learn in production. Improvements arrive as new versions through your change control.
Data integrity
ALCOA+ records, sha256 hash per frame, configurable retention, and exports ready for an inspector.
On-prem or air-gapped
Inference and records stay inside your validated network. Nothing leaves the site unless you decide it does.

Start a pilot

One station, one product, a few dozen good units.

Pilots run alongside your current inspection, so you see what Phamqyra catches — and what it stops rejecting — before anything changes in your batch record.

Register See pricing