01
MIND
A full-stack neuroscience and psychology exploration app built to answer “why does my mind do this?” rather than give generic advice. Conversations run through a modular AI pipeline with RAG, a non-overridable safety gate, and citations backed only by real DOI/PMID papers.
Overview
MIND is a greenfield full-stack product designed as an evidence-based alternative to generic AI mental-health chatbots. The core product question is “Why does my mind behave this way?” — not “What should I do?” It is an educational platform for exploring neuroscience and psychology through structured, cited conversations.
Problem
Typical chatbots invent citations, collapse every query into generic advice, and treat scientific knowledge and personal memory as the same context. MIND needed a system that could retrieve real papers, keep personal data separate, and refuse unsafe or overconfident responses before they reach the user.
Approach
Built a multi-agent conversation pipeline instead of a single LLM prompt. Each turn flows through intent analysis, context, psychological patterns, knowledge retrieval, evidence synthesis, neuroscience and psychology reasoning, intervention design, a non-overridable safety gate, then the response. Scientific knowledge and personal memory are stored separately.
Architecture
The backend is a FastAPI service with modular agents. Retrieval uses PostgreSQL with pgvector; research integrates PubMed and Semantic Scholar. Citations are never invented — only papers with verified DOI or PMID identifiers are shown. Next.js 15 delivers a public site plus an authenticated workspace with streaming SSE replies. Auth is production-grade: Argon2id passwords, opaque hashed sessions, TOTP MFA, OAuth scaffolding, rate limits, audit logs, and IDOR-safe ownership checks. Conversation bodies are not logged. Deployed as a split architecture: Vercel for the app, Render for the Dockerized API, and Neon for Postgres, with a Next.js /api proxy handling SSE and first-party session cookies.
Product surface
The app includes a research explorer, thought and experiment tools, insights, and a full account system with MFA, session management, and privacy controls for exporting or deleting data.
Result
A live educational platform with streaming conversations, verified citations, and a privacy-first backend. The pipeline stays inspectable: retrieval, reasoning, and safety are separate stages rather than one opaque prompt.
Technology
Next.js 15 · React 19 · TypeScript · Tailwind · FastAPI · Python · PostgreSQL · pgvector · Redis · Groq (Llama 3.3) · SSE · Docker · Vercel · Render · Neon