Why these repositories?

Two existing CEDA repositories as first candidates for SDP deployment:

Both already work locally and have decorators and config for usage with S3 / SQL.
Both are also packages and can thus be deployed together via an independent pipeline or integration repository.

Caspar and Tomer (CEDA) have experience with these repositories.

Gaps these repos naturally surface:

1cijferho — Overview

Converts fixed-width DUO files (1CijferHO format) into analysis-ready CSV/Parquet.

Key characteristics:

• 7-step sequential pipeline — each step independently callable
• Streamlit UI — wizard-like interface for non-technical users
• PII handling — automatic SHA256 hashing of BSN, Persoonsgebonden nummer
• Multiprocessing — parallel chunk-based conversion for large files
• Config-driven — file type detection via prefix patterns, metadata-driven field mapping

1cijferho — Architecture

Dependencies: polars, pandas, streamlit, pyjanitor, rich, fastexcel

1cijferho — Relevance

Notes:

• Streamlit, S3 and SURFconext are already tested on SDP by Caspar
• 1cijferho processes PII (BSN, Persoonsgebonden nummer) — deploying it is a good opportunity to jointly explore how PII handling on SDP could work (GAP-01)
• The 7-step pipeline is a concrete case for data lineage tracking (GAP-04)

NFWA — Overview

No Fairness Without Awareness — R package for analyzing bias in student data.

Developed by LTA lectorate (The Hague University). Trains ML models to predict student retention, then evaluates fairness across sensitive variables (gender, prior education, socioeconomic status).

Pipeline:

NFWA — ML components

NFWA — Architecture

Dependencies: tidymodels, ranger, glmnet, DALEX, fairmodels, ggplot2, quarto

NFWA — SDP relevance

Test questions:

• MLflow integration effort for tidymodels workflow? (GAP-05)
• Model versioning — what metadata needed for AI Act Art. 9/13?
• Fairness metrics as structured data — feed into monitoring?
• ML predictions on student retention need auditable decision logs (GAP-03)

Both repos demonstrate the matryoshka pattern

Config switch determines where each level connects:

• Local: filesystem, console output, no auth
• SDP: S3/PostgreSQL, Loki logging, SURFconext auth

What we want to learn

These repos surface 4 of the 14 gaps identified in the SDP gap analysis — practical test cases instead of abstract requirements.

Collaboration with team SDA

Caspar and Tomer have hands-on experience with both repositories.

Joint experiments (Sep–Dec 2026):

• Deploy 1cijferho on SDP — surfaces PII handling and lineage questions
• Deploy NFWA as batch job — surfaces model registry and audit logging questions
• Evaluate possible tooling (MLflow, OpenLineage) with real workloads
• Measure operational overhead and document findings

Timeline

Findings feed into tooling choices and gap resolution for the broader Educational Data Space platform.