Skip to content
GAIAA, Intelligence, engineered
Solutions
Solutions overviewBuild, deploy and improve AI systems that deliver real value.Private AI infrastructureRun private AI securely in your infrastructure.Existing AI optimizationMake existing AI faster, more reliable and more valuable.
Main specializations
Computer VisionSmart CityLIMSIndustrial IoT
IndustriesCase StudiesAbout usBlogContact
ENPL
Solutions
Solutions overviewPrivate AI infrastructureExisting AI optimization
Main specializations
Computer VisionSmart CityLIMSIndustrial IoT
IndustriesCase StudiesAbout usBlogContact
ENPL
GAIAA, Intelligence, engineered

Your signals, systems and constraints become the starting point for AI that can work in the real world.

Solutions overviewIndustriesCase StudiesAbout usBlogContact

Bring a process, signal or recurring decision.

Start the conversation
GAIAA sp. z o.o. sp.k.
ul. Sofoklesa 17
60-461 Poznań, Poland
NIP: 7811993317
KRS: 0000777888
© 2026 GAIAAPrivacy policyhello@gaiaa.eu+48 666 308 008LinkedIn
ENPL

Case study: Laboratories & science

Bring compute to governed scientific data without breaking lineage.

Versioned local runs keep governed data, code, runtime, results and review state connected.

SystemOutcomesRelated workDiscuss your case

System

The project and the problem

Researchers in the laboratory programme needed Python, notebooks, machine-learning models and graphics-processor workloads for confidential measurement data that had to remain on local infrastructure. A conventional shared server would create dependency conflicts, uncontrolled resource use and result files whose input dataset, code, runtime and model versions could not be reconstructed.

How the process worked

A control plane recorded the user, project, governed measurement inputs, code version, runtime and resource limits for each request. Isolated workers executed the CPU or GPU job locally; immutable artifacts and results were then validated and published as separate Analysis records linked back to their exact inputs.

Auditable local run

Advanced analysis, without moving governed data.

Use the controls below to inspect the key stages, constraints and validation decisions in that process.

Auditable local run

RUN-2478 · Coating uniformity predictor

Recorded run · select any stage to inspect its evidence.

REVIEW READY

Compute zone
LAB-COMPUTE-01 · on premises
Egress policy
Deny
Observed external egress
0 B
Scheduler job
JOB-98341
Governed input

DS-1842 · ALD thickness study

Population
1,248 approved observations · 12 batches
Snapshot
30 sie 2026 · 15:06
Schema
Thickness study v3
Snapshot digest
SHA-256 61d8…9a2f
Analysis definition

Uniformity predictor · candidate v7

Git commit
8ac41e
OCI image digest
sha256:2b19…a441
Reproducibility
Seed 42 · deterministic mode recorded
Compute
LAB-GPU-02 · 48 GB VRAM
Execution evidence

From snapshot to review bundle

Selected evidence

Evaluate batch holdout

Three batches excluded from training met the approved error criterion.

Holdout
3 batches excluded from training
MAE
0.83 nm
Criterion
≤ 1.00 nm · pass
Scientific result

Batch-held-out result

CRITERION PASSED
Holdout result for RUN-2478
Measured versus predicted coating thicknessRUN-2478 holdout result. Eighteen observations from three batches excluded from training are plotted against an identity line and eight nanometre tolerance band. Mean absolute error 0.83 nm. Approved criterion ≤ 1.00 nm · pass. Review bundle ART-2478.8080120120160160200200Measured thickness (nm)Predicted thickness (nm)
Holdout MAE
0.83 nm
Criterion
≤ 1.00 nm · pass
Holdout
3 batches excluded from training
Review bundle
ART-2478
View holdout values
ObservationBatchMeasuredPredictedResidual
OBS-01B-1094.0 nm93.0 nm-1.0 nm
OBS-02B-10101.0 nm101.6 nm0.6 nm
OBS-03B-10108.0 nm106.9 nm-1.1 nm
OBS-04B-10116.0 nm116.7 nm0.7 nm
OBS-05B-10121.0 nm120.5 nm-0.5 nm
OBS-06B-10127.0 nm127.8 nm0.8 nm
OBS-07B-11130.0 nm131.2 nm1.2 nm
OBS-08B-11138.0 nm137.1 nm-0.9 nm
OBS-09B-11142.0 nm142.4 nm0.4 nm
OBS-10B-11148.0 nm149.0 nm1.0 nm
OBS-11B-11151.0 nm150.3 nm-0.7 nm
OBS-12B-11158.0 nm158.8 nm0.8 nm
OBS-13B-12164.0 nm162.7 nm-1.3 nm
OBS-14B-12167.0 nm167.5 nm0.5 nm
OBS-15B-12173.0 nm173.9 nm0.9 nm
OBS-16B-12179.0 nm178.4 nm-0.6 nm
OBS-17B-12184.0 nm185.1 nm1.1 nm
OBS-18B-12190.0 nm189.2 nm-0.8 nm
  1. ✓CandidateRecorded
  2. ✓Automated validationPassed
  3. ○Scientific reviewReady
  4. ⌁DeploymentLocked

Automated checks passed. Human review is still required.

Input
DS-1842 · immutable
Code
8ac41e
Environment
2b19…a441
Seed
42 · deterministic
Result
ART-2478 · immutable

This result resolves to DS-1842, Git 8ac41e, image 2b19…a441 and JOB-98341.

Outcomes

Operating constraints

  1. Controlled local boundary
  2. CPU and GPU workloads
  3. Versioned code and models
  4. Immutable result artifacts

What changed?

  1. Reconstructable local runs
  2. Measurement-to-result lineage
  3. Controlled CPU and GPU execution
  4. Review-gated model release

Evidence retained: Each run retained its governed inputs, code version, runtime, resource request, artifacts and review state.

Related work

SolutionPrivate AI infrastructureSolutionLIMSSolutionExisting AI optimization

Discuss your case

Have a similar operating problem? Share the environment, constraint and result you need. We can discuss the engineering path without exposing confidential details.

Discuss a similar challenge