Needs assessment pipeline

Each partner had adapted the core questionnaire locally. Response options had drifted, district names were spelled three ways, and household size was a number in some forms and a bracket in others.

Cluster reporting deadlines did not move, so the merge was done under time pressure every round — which is exactly when transcription errors get introduced.

I wrote a canonical schema for the assessment and a per-partner mapping into it, so each agency could keep the form its enumerators were trained on.

Geographies were resolved against an official administrative gazetteer rather than free text, which made district-level aggregation reliable for the first time.

The pipeline runs on ingest and produces both the cluster submission format and a partner-level quality report showing what failed and why.

Partners kept their forms, the consortium got one dataset, and the quality report moved the conversation from arguing about whose numbers were right to fixing the specific records that were wrong.