- Where
- USFca Professors: Alark Joshi, Kelly L'Engle
- Team
- 3-person team
- My role
- backend and AI backend, RAG pipeline, database design
- When
- 2025-08 - 2026-07
Problem
Structured wellness coaching depends on a human coach's judgment, which does not scale to everyone who needs it. A general-purpose chatbot can hold the conversation but has no grounding in what a trained coach would actually say at a given point in a program.
Solution
A RAG pipeline that retrieves relevant examples from a corpus of real coaching interactions and conditions the model on them before it answers, so responses stay anchored to evidence-based practice rather than to the model's priors.
Architecture
Overview
Nala is an intelligent health coaching assistant that guides users through a structured 4-week wellness program for stress management and habit building. It combines conversational AI with evidence-based coaching via a three-layer architecture: a React Native mobile frontend, a FastAPI backend with SQLAlchemy and PostgreSQL, and a RAG (Retrieval-Augmented Generation) pipeline that retrieves relevant coaching examples from a pgvector database before generating contextually appropriate responses through Claude or GPT. Built on real interactions from the Examen Tu Salud program. Currently under active maintenance with a coauthored publication submitted to CHI 2026.
Impact
- Coauthored publication entry submitted to ACM CHI 2026.
- Built on real interactions from the Examen Tu Salud program.
- User testing on real students and faculty.