Traceable experimental inputs
Provenance, licensing, split policy, preprocessing, class balance, seeds, and dataset limitations.
Supporting project · Scientific ML engineering
Evaluation hardening plannedProduct Image Classifier
A small-data image classification project being reframed around reproducible experiments, calibrated decisions, explicit limitations, and portable inference.
Engineering thesis
An ML metric is not evidence by itself.
A defensible result requires versioned data provenance, leakage-resistant splits, controlled preprocessing, seeded experiments, meaningful baselines, uncertainty, calibration, and error slices. Training and inference artifacts must remain separate and traceable.
The project is accurately scoped as classification; it does not claim object localization unless localization behavior and evaluation are implemented.
Target evidence
What the evaluation must prove
Defensible comparison
Baselines, confidence intervals, calibration, per-class errors, ablations, and reproducible commands.
Versioned model delivery
Model and data cards, checksums, environment lock, lineage, and a distinct inference package.
CPU-first inference
A validated CPU path plus separately measured GPU support without conflating serving and training portability.
Current boundary
Historical results require revalidation.
- No accuracy, calibration, latency, or portability claim is promoted until its methodology and reproduction path are published.
- Repository modernization is planned; current public metadata states that status explicitly.
- The release goal is scientific clarity, not added infrastructure for appearance.
Related engineering evidence
Connect evaluation to delivery.
AI Prompt Optimizer
A shipped AI system with a versioned evaluation corpus, deterministic fallback, and explicit claim boundaries.
All engineering case studies
Compare scientific ML evidence with product, platform, infrastructure, and backend reliability work.