Skip to content
GROUNDED HEALTH AI · CLIENT Yuktha Wellness

Yuktha Wellness

Paid multilingual health assistant: hybrid retrieval, reranking and an emergency check that runs before any answer.

Illustration of Yuktha Wellness: a web-chat answer with a cleared emergency check beside the emergency-first retrieval pipeline
ILLUSTRATION
CURRENT SCOPE

Paid client work · production assistant

WHO IT'S FOR

Client project · Yuktha Wellness

Women's health startup in India. Paid milestone work on their production assistant, web chat and WhatsApp.

THE ENGINEERING

Paid client engineering for a multilingual health assistant: hybrid dense and keyword retrieval, cross-encoder reranking, emergency/crisis handling, structured grounding checks, SSE streaming and Redis caching. Web chat and controlled WhatsApp routing connect to the platform. Later work separates user-authored memory from generated replies and improves multilingual retrieval and image-only handling. The August release record includes 28 passing backend test files and a nine-check namespace migration. The recorded full latency target remains unmet; clinical content and final tone require human review. Historical deployment drift remains documented, so these records do not establish current source-to-live parity.

Node.jsHybrid RAGPinecone

BEFORE → AFTER

Real problems from this build, what I changed and the verified result.

  1. CASE 01
    Before

    Multilingual knowledge assistants need relevant evidence rather than similarity scores alone.

    After

    Retrieval behavior was exercised through recorded multilingual evaluation and backend checks.

    How I fixed it

    Combined dense and keyword search with cross-encoder reranking and explicit grounding checks.

  2. CASE 02
    Before

    Healthcare conversations need a separate emergency and mental-health crisis path.

    After

    Recorded checks cover appropriate crisis responses; clinical content still requires qualified human review.

    How I fixed it

    Placed deterministic crisis detection before retrieval and generation, including transliterated language cases.

  3. CASE 03
    Before

    User memory must distinguish personal statements from generated assistant text.

    After

    Memory provenance is handled explicitly; recorded full latency targets remain unmet.

    How I fixed it

    Separated user-authored facts from generated responses and excluded legacy mixed records.

Evidence reviewed: August release records · October 10 public case study