za.
All work

MACHINE LEARNING / APPLIED AI

MyCareAI

An emotion-aware AI assistant combining text classification, facial emotion recognition, and adaptive chatbot responses.

Period
June 2025 capstone; research extension documented in 2026
Status
Prototype
Original MyCareAI project logo and tagline

Original MyCareAI project logo and tagline

The problem

An assistant that responds to emotional context needs to connect model predictions with the user-facing conversation.

Contribution & scope

An individual role is not specified in the available project material.

Project implementation

  • Separate chat, text-classification, and facial-emotion services behind an Express API.
  • A four-class thesis text classifier covering sadness, happy, anger, and fear.
  • Optional TFLite / MediaPipe facial-emotion context and a persona-preview endpoint that works without an LLM call.
  • Facial-emotion experiments documented with CNN/VGG16 and FER-2013 / RAF-DB, alongside the thesis repository's optional FER service.
  • Docker Compose orchestration for the Vite frontend, API, and both Python services in the thesis prototype.

The approach

The original capstone connected a web client to a Node.js / Express gateway, a Flask / TensorFlow text-emotion classifier, and Vertex AI. The thesis extension uses Gemini API for chat and adds optional camera-based facial emotion recognition. An emotion-aware context manager selects an adaptive response persona using text, facial signals, and risk rules.

Architecture & workflow

  1. Text / optional camera input
  2. Preprocessing
  3. Emotion services
  4. Context + risk rules
  5. Gemini chat
  6. Adaptive response

Technical challenges

  • Connecting text and facial emotion predictions to a coherent chatbot experience.
  • Separating model inference from the main application service.

Outcomes

  • Connected emotion inference with generative chat through separate application services.
  • Extended the capstone into a thesis prototype with adaptive personas and optional facial context.

Evidence & context

  • Connected a TensorFlow emotion classifier, Flask inference service, Node.js gateway, and Vertex AI chatbot.

    Architecture documented in the public project README. The application is an emotion-aware prototype, not validated mental-health care.

    MyCareAI README — architecture and technology

  • Extended the application with Gemini chat, optional facial-emotion context, and adaptive persona selection.

    Documented in the public thesis repository and its service code. This establishes implementation scope, not clinical effectiveness or independently measured model accuracy.

    MyCareAI thesis repository

Scope & limitations

  • The available evidence does not provide a reproducible text-model evaluation protocol; numerical accuracy claims are withheld from this public case study.
  • This emotion-aware prototype has not been clinically validated and does not establish a medical diagnosis.

What I learned

  • Model outputs become useful when they inform a clear interaction in the product.
  • Separate inference services allow model development and application integration to progress independently.

Built with

  • Python
  • TensorFlow
  • Flask
  • Node.js / Express
  • TFLite / MediaPipe
  • Gemini API
  • Docker Compose