Data Beat

Home Screen
Contradictions
Home Screen
Add comments
Evidence of Notebook run

🚀 Data Beat — Turning messy healthcare data into decisions you can trust Inspiration: Healthcare planning relies on data that looks complete but isn’t trustworthy. We saw thousands of facility records with contradictions, vague claims, and inconsistent locations—and realised: the real problem isn’t missing data, it’s unreliable data.

What it does: Data Beat turns messy healthcare facility data into transparent, reviewable, decision-ready insights. The final Agent output that classifies data into contradictions is only used against 100 records as we ran out of SQL credits on free version.

Detects contradictions across location, capabilities, and text Extracts real capabilities from noisy descriptions Shows clear evidence for every claim Scores data quality and risk Prioritises records for human review Lets non-technical users approve, fix, or override data

How we built it: Databricks Lakehouse (Delta) for data pipelines AI Agent + Model Serving for contradiction detection Custom ontology + tagging system for standardisation Hybrid approach: ✅ Rule-based checks (geo, numeric) ✅ LLM-based reasoning (text vs structured data) Databricks App for review workflow Full audit trail: every score, claim, and decision tied to evidence

Challenges: Messy, inconsistent text with marketing claims Structured fields and descriptions often contradicted each other Ensuring LLM outputs were reliable, structured, and explainable

What we’re proud of ✅ Built a fully explainable AI system (no black box) with proper ontologies ✅ Created a 3-class contradiction framework (Hard, Logical, Weak Evidence) ✅ Designed a system where humans + AI work together ✅ Delivered a tool that non-technical users can actually trust

What we learned: Trust > volume in data systems AI must show evidence + uncertainty, not just answers Ontologies are essential for scaling messy real-world data Human-in-the-loop is critical for real decision-making

What’s next for Data Beat: Learn from reviewer feedback to improve accuracy Add district-level demand insights (NFHS) Detect duplicates and merge entities Expand beyond India → global healthcare planning