← Back to Projects

Convo Care — Privacy-First AI Risk Triage Platform

Designed and built a privacy-first AI system that analyzes text and voice conversations to produce sentiment and mental-health risk scores without storing transcripts or raw audio.

Role: System Architect & Lead Engineer Tags: ai-systems · privacy · nlp · full-stack

Highlights

  • Built a full-stack platform (FastAPI + React) for text and voice-based risk scoring
  • Implemented in-memory speech-to-text and NLP scoring with zero transcript persistence
  • Designed a scores-only data model for privacy-preserving analytics and trend tracking
  • Delivered per-student risk trends, history, and high-risk flagging

Impact

  • Demonstrated how AI systems can be designed with privacy as a first-class constraint
  • Produced a reusable pattern for compliant AI decision-support pipelines
  • Showcased end-to-end system design: ingestion, scoring, storage, and visualization

Context

In sensitive domains like mental health, data privacy is a core system constraint. Storing transcripts or audio creates unnecessary risk, even when the goal is only trend analysis and triage.

This project explores a scores-only AI pipeline:

Analyze → score → discard raw data → retain only safe metrics.


What This Is

A web-based AI decision-support platform that:

  • Accepts text or voice input
  • Transcribes audio in memory
  • Runs sentiment and risk scoring models
  • Stores only numerical scores and safe metadata
  • Presents trends, history, and high-risk flags in a React dashboard

The backend is built with FastAPI and MySQL, and the frontend uses React with WebRTC for voice capture.


Outcomes

  • A working demonstration of privacy-first AI system design
  • A full-stack, containerized platform with real-time scoring flows
  • A reusable architectural pattern for compliant AI analytics systems

Why This Matters

This project reflects how I design AI in production environments:

Treat data handling, privacy, and compliance as system-level architecture decisions, not afterthoughts.