National Census Digital Assistant (HPE)

A civic-scale conversational assistant supporting India's census operations, built at HPE on Azure. It grounds guideline and FAQ answers with retrieval-augmented generation, and runs National Population Register enrolment and birth and death reporting as deterministic, validated state-machine flows against central government registries. Every backend call sits behind Azure API Management as a managed, audited boundary.

I worked against a government program's compliance and audit requirements. I designed the conversational architecture around one split that keeps it safe: generative answering handles open questions, and deterministic transactional flows handle anything that writes a civic record. I owned the guardrail and audit posture, so the assistant never speculates on policy and every registration step stays traceable.

1. Deterministic state machines for registrations over end-to-end generative flows. Civic records demand validation and receipts, not fluency.
2. Azure API Management as a hard boundary, so government-data access is a managed, auditable contract rather than a direct dependency.
3. Guardrails and explicit refusal paths, so the assistant answers what it can ground and declines what it cannot, instead of inventing civic policy.

At civic scale, one wrong answer about enrolment or a mishandled birth or death record is a real-world failure. The architecture keeps the parts that touch records boring and deterministic. The parts that talk to citizens stay helpful, inside firm bounds.