Limited labelled documents make it difficult to cover varied categories, layouts and quality needs. Naive generation can simply amplify bias or duplicate the generator's assumptions.
Shivanand Kumar← Back to public work
Independent build · Synthetic data · CI passing
Synthetic data · Decision support
Synthetic Data and Print Recommendation Agent
An evaluation-first project for a data-scarcity problem: create useful variation without treating more synthetic rows as proof of better recommendations.
12synthetic holdout cases
10×best evaluated scale
100×saturation detected
Humanreview-required flag
Problem and ownership
A real engineering question, built to be inspected.
I built the scarcity analysis, controlled generator, feature pipeline, local recommender, API, synthetic holdout evaluation, safety rules, tests and CI workflow.
Verified output
Concrete behavior, not a concept slide.
Abbreviated verified response shape from the brochure test. Separate policy-boundary tests confirm that ambiguous inputs are held for human review and printing is never executed.
{
"document_id": "incoming_brochure_001",
"recommended_settings": {
"color_mode": "color",
"quality_mode": "high"
},
"evidence_document_ids": ["…"],
"confidence": "scored",
"execution_performed": false
}What the repository proves
Designed, tested and documented end to end.
- Coverage and rare-segment scarcity analysis
- Deterministic controlled generation at 1x, 10x and 100x
- Document feature engineering and structured recommendations
- Synthetic holdout evaluation isolated by document ID
- Diversity, duplicate, rare-group and saturation checks
- Human review for low-confidence and boundary cases
Architecture flowDesign decisions ↗
- 01
Measure gaps in the small synthetic seed corpus
- 02
Generate controlled variations across missing document strata
- 03
Engineer bounded document and quality features
- 04
Recommend structured settings with evidence and confidence
- 05
Compare scales on a separately versioned synthetic holdout with disjoint IDs
Limits and next validation
What this project does not prove—yet.
- The 12 holdout records use disjoint IDs, but their expected settings come from the same documented synthetic policy world. Results test controlled scale and saturation, not out-of-policy or expert-labelled generalization.
- Synthetic holdout performance does not establish real-world print quality or business outcomes.
- The service uses structured document metadata rather than raw PDF rendering or printer telemetry.
- A real deployment would require expert-labelled data, printer constraints, user feedback and drift monitoring.
Credibility boundary
Independent work, described honestly.
Every document profile and label is synthetic. The repository contains no employer documents, print rules, source code or confidential architecture, and the service cannot operate a printer.