TalkRx: Real-Time AI Anxiety Companion

Streamlit Application with Preloaded Conversations from Kaggle Data
Streamlit Application with OpenAI Agent Response for Anxiety Profiling from Input Conversation
TalkRx Flutter app chat screen view showing conversation between the agent and user
TalkRx Flutter app home page showing future options to enhance application functionality

TalkRx – AI‑Guided Anxiety Support Built for Responsible Healthcare

1. Inspiration

An estimated 1 in 5 adults in the U.S. experience an anxiety disorder each year, yet more than half don’t receive treatment. We noticed two major gaps contributing to that. First, support isn’t always available when people need it most. Therapy requires time, money, and scheduling, which many people can't access on demand. Second, while AI chatbots are widely available, they often lack clinical accuracy, personalization, and trust. Some people also aren’t ready or willing to talk to a therapist directly due to stigma, fear, or just not knowing where to start. We wanted to build something that bridges all those gaps, that's more reliable than Googling symptoms, and more accessible than a therapy session. Our goal was to deliver real-time, evidence-based anxiety support that people can trust and actually use when they need it most.

2. What It Does

TalkRx is a cross‑platform mobile system where anxiety patients chat with an AI companion that can:
1. Extract an AnxietyProfile – symptoms, triggers, emotions, evidence, and a provisional DSM‑5‑TR / ICD‑10 diagnosis.
2. Suggest evidence‑based next steps – psycho‑education, coping tips, and FDA‑approved (or off‑label) medication options.
3. Share structured data with providers through FHIR v5 Patient + Observation resources, making human follow‑up fast and safe.

3. How We Built It

Architecture Diagram

Architecture Details

Area	Key Work
Infrastructure	FastAPI microservices with LangServe + Model Context Protocol (MCP) for tool discovery; PostgreSQL & DuckDB (structured data), MongoDB (JSON chat logs); Docker‑ized on an SSDNodes VPS.
AI Layer	GPT‑4o via Langchain routes; custom AnxietyExtractor parses chat text into a Pydantic schema and stores it.
Clinical Layer	Curated DSM‑5‑TR ↔ ICD‑10 mappings and a medication knowledge‑base (doses, interactions, black‑box warnings).
App Layer	Flutter/Dart UI: Home + Chat dashboards consume REST endpoints; Streamlit evaluator lets clinicians review AI output and tune tone/safety.
Standards	FHIR v5 resources generated with fhirstarter; MCP enables the agent to call any microservice like a plug‑in.

Team Roles

Rishov – backend, AnxietyExtractor, databases, FHIR, MCP. Usma – clinical rules, DSM/ICD mapping, medication DB, output validation. Mohith & Prakriti – Flutter UI, state management, API wiring with Dart. Miao – data persistence & privacy controls using Python, gRPC, and FastAPI

4. Core Features Delivered

Agent Chat Service – OpenAI + LangServe endpoints.
Patient Data & Anxiety Extractor Service – exposes MCP tools.
AnxietyProfile extractor (Langchain‑powered).
Clin‑grade tables: disorders + meds in Postgres via SQLAlchemy ETL.
Flutter client (Home & Chat).
Evaluator Web App – Streamlit tabs to view Kaggle chat data (DuckDB) and live‑test agent replies.
One‑click VPS Deploy – all three services shipped in Docker.

5. Challenges & Solutions

Scarce, clean mental‑health data → curated Kaggle conversations and hand‑labeled ICD‑10 examples.
Empathy vs. accuracy → iterative tone tuning with evaluator tool.
Hackathon time limit → feature flags: chat first, then meds, FHIR, gRPC/MCP.
Dataset approvals (e.g., MIMIC‑IV) were slow → pivoted to public corpora from Kaggle mental healthcare datasets.

6. Accomplishments

End‑to‑end pipeline: raw text → structured AnxietyProfile → tailored recommendations.
1,200‑entry medication DB with dosing, interactions, & special‑population notes.
Seamless MCP mounting so GPT‑4o “discovers” any TalkRx tool without code changes.
Flutter app live‑tests new patient chats and streams AnxietyProfiles instantly.

7. What We Learned

Backend: mounting MCP servers inside FastAPI, generating FHIR v5 on the fly.
Clinical: programmatic DSM‑5‑TR translation and scalable AI output QA.
Frontend: Flutter for healthcare, async state, AI microservice orchestration.
Full‑stack skills: privacy, logging, Docker deploy under 48 h.

8. Next Steps

Ingest larger validated sets (e.g., DAIC‑WOZ) for model fine‑tuning.
FHIR push to EHR sandboxes; pilot with university counseling centers.
Expand user wellness: Calm‑style audio, journaling, peer groups.
Pitch to VCs: reduce no‑shows, speed triage, inject structured data back into provider workflows.