~/kaustubh-lall

~/projects $ cat projects/dusk.md

Dusk

privateproprietary

ResMed · San Diego, CA

A virtual-therapist conversational agent supporting patients waiting to receive care, built with RAG, persistent memory, multi-agent architecture, and voice-to-voice interaction.

Dusk was a virtual therapist and patient-support platform designed to help patients while they waited to receive care. I built it at ResMed. It combined retrieval-augmented generation, custom persistent and user-specific memories, an agent-based architecture with specialized agents, and full voice-to-voice interaction (speech-to-text and text-to-speech) over a complete supporting backend. It developed personalized goals and targets based on CBT-I guidance, classified patient phenotypes, and adapted its guidance to the classified phenotype.

Technologies

  • LangChain
  • RAG
  • Speech-to-text
  • Text-to-speech
  • Python

Skills

  • Agent architecture
  • RAG
  • Conversational AI
  • Voice interfaces
  • Backend systems

What I built

  • Built Dusk from the ground up: the agent architecture, RAG pipeline, memory system, voice stack, and supporting backend.

What it did

Dusk acted as a virtual therapist and conversational companion for patients awaiting care: voice-to-voice conversations, guidance grounded in retrieval-augmented generation, and memory that persisted across sessions and adapted to the individual user.

It built personalized goals and targets from CBT-I (cognitive behavioral therapy for insomnia) guidance, classified each patient's phenotype, and adapted its recommendations to that phenotype.

Architecture (capability level)

An agent-based architecture with specialized agents for distinct responsibilities, retrieval-augmented generation over curated knowledge, custom persistent user-specific memory, LangChain orchestration, speech-to-text and text-to-speech for full voice interaction, and a complete supporting backend.

notes

  • Proprietary ResMed work, described through capability patterns only, without confidential implementation details.
  • Intellectual-property filing pending.
← All projects