Kalibra

A governed, reproducible visual-inspection runtime for industrial anomaly detection.

Kalibra is an offline visual-inspection project that carries a PaDiM anomaly signal from a governed VisA proxy dataset into ONNX Runtime, then verifies the integrated runtime with deterministic checks and hash-anchored evidence. The current evidence is single-seed and proxy-domain bounded; it does not demonstrate calibrated confidence, production readiness, or domain-of-record industrial performance.

Designed and built by André Ramos as an AI/ML engineering portfolio project.

What This Project Demonstrates

• Runtime equivalence verified across 6,492 samples: maximum absolute deviation 7.105e-15 against a 1e-12 tolerance (evidence).
• Deterministic replay passed all 7 governed comparisons for the documented runtime evidence chain (provider integration).
• The dataset-to-runtime path is governed through VisA provenance, PaDiM artifacts, ONNX export, runtime replay, and hash records.
• Runtime equivalence proves signal preservation through the integrated runtime; it does not prove accuracy, calibration, production readiness, or cross-domain validity.

Technology Stack

Area	Evidenced technologies
Runtime and evaluation	Python, NumPy, Pillow, PaDiM, ONNX, ONNX Runtime, pytest
Static portfolio	HTML, CSS, vanilla JavaScript

VisA is used as the governed proxy dataset, not as the industrial domain of record.

Quick Verification

Kalibra has three explicit verification levels. Level 1 is the public clean-clone contract; Levels 2 and 3 require separately acquired governed VisA data.

Level 1 — Clean-Clone Verification

Quick Verification works using tracked repository contents only:

git clone https://github.com/andrejr03/kalibra.git
cd kalibra
python3 -m pytest -q
python3 scripts/verify_public_clone.py
python3 scripts/build_portfolio_evidence_bundle.py --check

This checks the default committed-fixture tests, model and governed-artifact SHA-256 records, the public evidence manifest, and portfolio bundle drift. It does not perform governed runtime replay or scientific reproduction.

Level 2 — Governed Runtime Verification

The full runtime-equivalence replay requires the separately governed VisA archive and extracted dataset:

python3 scripts/verify_padim_runtime_equivalence.py verify
python3 scripts/verify_placeholder_retirement.py verify
python3 -m pytest -q -m governed_data

The dataset is not shipped in Git. That is an intentional licensing and governance boundary, not a missing repository file. Acquisition, expected paths, and integrity steps are documented in VisA Acquisition and Governance.

Level 3 — Full Scientific Reproduction

Full scientific reproduction is a separate workflow requiring governed dataset acquisition plus the documented training, inference, evaluation, export, and runtime environment. It is not a default clean-clone guarantee. Follow the scientific evaluation methodology after completing the governed acquisition instructions.

Complete governed replay evidence is documented in Runtime Equivalence and Runtime Provider Integration.

Current Status

Implemented and evidenced:

• Governed VisA proxy dataset, PaDiM baseline, and single-seed scientific evaluation.
• Governed ONNX export and runtime, including export equivalence, runtime equivalence, and deterministic replay.
• Canonical placeholder retirement and a static portfolio generated from governed artifacts.

Designed, not yet demonstrated:

• Calibrated confidence, accept / review / reject routing, abstention, drift assessment, and interactive human review.
• End-to-end validation, domain-of-record industrial performance, and multi-seed statistical characterisation.

Kalibra is an offline, batch-oriented, single-seed, VisA-proxy, non-production engineering artifact.

Portfolio Experience

The portfolio is a static presentation that renders committed governed artifacts; it performs no browser inference and has no runtime dependency on external services or data sources.

• Live portfolio: https://andrejr03.github.io/kalibra/
• Repository source: portfolio/

Overview of the static evidence presentation and project boundary.

Governed runtime-inspection case showing the real VisA-proxy input, presentation-only PaDiM anomaly overlay, native raw anomaly map, recorded localization, and non-calibrated raw anomaly measure.

Runtime-equivalence station showing sample count, tolerance, maximum deviation, and replay status.

Scientific-boundaries station showing measured results, limitations, and absent trust capabilities.

Engineering Domains

Domain	Status	Evidence or next step
Inspection	Implemented and evidenced	PaDiM baseline and governed ONNX Runtime inspection path.
Trust Qualification	Designed, not yet demonstrated	Future calibration, abstention, and accept / review / reject qualification.
Human Review	Designed, not yet demonstrated	Future review routing and operator-facing evidence workflow.
Evidence	Implemented and evidenced	Hash records, governed replay files, and committed portfolio evidence bundle.
Evaluation	Implemented for the current slice	Single-seed VisA-proxy metrics, export equivalence, runtime equivalence, and replay checks.

Repository Structure

Path	Purpose
src/	Runtime, inspection, evaluation, evidence, and domain modules.
scripts/	Governed training, export, verification, and evidence-bundle commands.
tests/	Unit, integration, runtime, provider, and artifact verification tests.
artifacts/	Tracked PaDiM, ONNX, runtime, equivalence, and replay artifacts.
data/	Governed VisA manifests, provenance, and derived evidence records.
docs/	Architecture, methodology, engineering notes, and evidence documents.
portfolio/	Deployable static portfolio generated from governed artifacts.
assets/	Portfolio screenshots, visual assets, and future workbench concept image.

Roadmap

All roadmap items are future work:

• Trust calibration and qualification.
• Accept / review / reject, abstention, and human-review routing.
• Drift assessment, multi-seed characterisation, and end-to-end validation.

Future Concept: Workbench Prototype (Not Implemented)

This design concept explores the intended review surface; the calibrated confidence, routing, drift, and human-review capabilities shown here are not implemented or evidenced in the current runtime.

View the future workbench concept

Concept only: the interface visualizes intended trust qualification and review workflows that are not implemented in the current runtime.

License

This project is licensed under the MIT License. See the LICENSE file for details.