Voice AI + retrieval + citations

Health Voice RAG Demo

Part of Karim Ali's ML engineering portfolio.

Voice AIHealth RAGLiveKitCited AnswersRubric Review

Recorded sessions

Replay recorded runs

Recorded from real agent sessions. Select one to open the full replay view.

Live event stream

Every run exposes ordered routing, retrieval, review, and answer events.

Rubric answer review

Generated answers are checked for grounding, citation support, safety, relevance, and voice fit.

Cited RAG output

Final responses include source cards, retrieval attempts, score, and review judgement.

A realtime voice/text health assistant that routes questions, retrieves sources, reviews answer quality with a rubric, and returns cited informational responses.

Built as a portfolio project for engineering review: observable events, async backend integration, replayable sessions, and a usable voice/text interface.

Informational only, not a diagnosis. The agent can ask follow-up questions or return an insufficient-information response when source evidence or answer review is weak.

What it demonstrates

Shows both product UX and backend event flow in one demo.

Routes health questions before retrieval or answer generation.

Recovers from weak search with quality review and query rewrite.

Uses a rubric-based answer review for grounding, citation support, safety, relevance, and voice conciseness.

Displays grounded answers with source cards and workflow status.

System path

Capture
agent_session, user_transcript
Voice or text input starts a LiveKit-backed agent session.
Route
question_routing
The router decides whether to search, ask, or handle directly.
Clarify
followup_question
Vague or missing context becomes a follow-up question.
Retrieve
query_embedding, source_retrieval
The query is embedded and matched against health sources.
Review sources
retrieval_decision, retrieval_quality_review
Weak evidence is judged before the system answers.
Rewrite / retry
query_rewrite, retrieval_attempt
Weak retrieval or blocked answers can trigger another attempt.
Generate
grounded_answer
Accepted sources produce a cited answer candidate.
Rubric review
answer_quality_review
The candidate is checked for grounding, citations, safety, and relevance.
Deliver
health_answer
The UI renders the final answer, follow-up, or insufficient-information response.