Dr. T

Inspiration

Traditional healthcare software is cold and transactional, but we were inspired by a real-life physician, Dr. T, who has been helping her patients with wits, wisdom, and kindness throughout their treatment processes. We built 'Dr. T' — an empathetic, Socratic clinical decision support platform and patient-first wellness guide to carry her legacy of compassionate healthcare into a modern digital experience.

What it does Dr. T is an empathetic, voice-synthesized AI companion and multi-agent health dashboard styled with the soothing tone of a "Maternal Soulmate." It is designed to act as a loving emotional anchor while offering professional-grade clinical oversight, cognitive modeling, and life planning. Key capabilities include:

Qwen-2.5 Semantic Extractor & Relational Life Graph: Users can input raw, unstructured conversation logs (e.g., "My mother Jane loves hot chocolate, and I take red pills before bed"). Dr. T analyzes the transcript, extracts key lifestyle facts, scores their semantic saliency (strength), proposes connections to existing memories, and dynamically maps them to an interactive relational graph.

Active Context Frame: These extracted nodes are held in active memory. To protect the LLM’s token context budget and reflect realistic memory processes, unpinned memory nodes undergo gradual mathematical decay unless pinned/locked by the user.

Agent Swarm & Socratic Voice Intercom: When faced with complex multi-disciplinary problems, Dr. T delegates work to a team of specialized sub-agents (e.g., Medical, Educational, Financial, and Legal) who run role-decomposed analysis in parallel, resolve conflicting goals, and feed their findings back to Dr. T for a warm, unified maternal Socratic synthesis.

How we built it We built Dr. T as a high-fidelity, full-stack application leveraging modern web and AI engineering frameworks:

Frontend: Implemented a responsive, desktop-first SPA in React 18 with Vite and styled using customized Tailwind CSS. Lucide-React supplies clean, modern icons, and SVG-rendered responsive links draw real-time connections on our interactive Memory Graph Canvas.

Backend Server: An Express.js server acts as our API Gateway, hosting endpoints for secure, server-side AI processing. It also serves Vite-compiled assets under production environments.

AI Intelligence (Gemini + Qwen Heuristics): Integrated the server-side @google/genai TypeScript SDK to analyze unstructured transcripts. It classifies thoughts into six strict categories (family, preference, health, learning, career, landmark), scores their cognitive relevance, and outputs clean structured JSON matching our node schemas. A robust client-side semantic fallback handler ensures graceful, zero-downtime execution if external rate limits are met.

Voice Synthesis: Implemented a responsive web speech integration supporting custom vocal avatars to bring Dr. T and her sub-agents to life during live, verbal consultations.

Challenges we ran into State Alignment with Dynamic Graph Edges: Connecting newly extracted nodes with existing nodes required maintaining strict state integrity across React state renders. We solved this by implementing stateful bidirectional connection mapping directly within the parent workspace, which automatically updates parent nodes when a child link is established.

API Resilience & Context Overload: Heavy, multi-agent conversational workloads are prone to latency and API token exhaustion. We resolved this through two methods:

Establishing a local semantic parser on the frontend that seamlessly steps in as a high-fidelity fallback.

Implementing a custom Memory Decay & Compact Graph Engine that simulates forgetting, allowing users to prune decaying, stale memories to preserve the agent's long-term focus.

Accomplishments that we're proud of Real Semantic Splicing: Designing a system that successfully bridges natural conversational entries with structured, interactive nodes without relying on hardcoded mock values. Seeing the graph update in real-time with precise strength parameters is incredibly rewarding.

Voice-Enabled Socratic Dialogue: Creating the Socratic Voice Intercom which synthesizes a realistic, multi-turn verbal dialogue between Dr. T and her specialized sub-agents. It feels like stepping into a live, collaborative clinical board meeting.

Clean Design System: Achieving a balanced visual hierarchy that pairs high-contrast slate cards with elegant display typography, smooth micro-animations, and responsive layouts.

What we learned Role Decomposition Is Vital: Rather than asking a single LLM to play multiple roles simultaneously, splitting the pipeline into targeted sub-agents (e.g., specialist nodes) yields much higher decision accuracy and clearer analytical boundaries.

The Power of Graceful Degradation: Designing software to fail gracefully—utilizing client-side heuristic fallbacks—is what separates fragile, proof-of-concept AI apps from stable, production-ready interfaces.

What's next for Dr. T Durable Cloud Synchronization: Integrating a persistent Firestore backend to synchronize the relational memory graph and active checklists securely across multiple devices.

Passive Ambient Extraction: Transitioning from manual chat/transcript entries to passive, real-time voice streaming. Dr. T would gently listen to daily ambient conversations and automatically update her memory graph in the background without needing manual inputs.

Advanced Biometric Integrations: Syncing with wearables (such as Fitbit or Apple Health) to feed live heart rate variability (HRV) and sleep parameters directly into the Biomedical Suite for deeper diagnostic modeling.

Built With

• all