Prach AI: Technical Architecture and Development Plan

This document outlines the technical architecture and development plan for Prach AI, a platform designed to empower the neurodiverse community.

I. User Interaction & LLM

• Multimodal Input: Supporting diverse modalities like speech, touch, and gestures for seamless user interaction.
• Predictive LLMs: Utilizing Large Language Models (LLMs) to predict user responses, powering Automatic Speech Recognition (ASR) and interpreting visual cues for dynamic communication.

II. Content & Gamification

• Gen AI Content: Generating engaging learning content (images, videos) for children using generative AI models to visually explain concepts.
• Gamification: Integrating gamification with Dash Plotly for displaying scores and progress to motivate and encourage user engagement.

III. Model Enhancement & Data

• Mixture of Experts (MoE) Approach: Employing a Mixture of Experts approach, combining Gemini with LLAMA for better, more nuanced responses.
• Data Grounding: Prioritizing data grounding by creating our own preferred content to combat misinformation and ensure trustworthy information.

IV. Accessibility & UX

• Neurodiversity-Friendly Prompts: Conducting extensive prompt engineering to craft neurodiversity-friendly responses tailored to individual needs.
• WCAG Compliance: Incorporating Web Content Accessibility Guidelines (WCAG) into our UX design for an inclusive and user-friendly experience.

V. Specialized SLMs

• Targeted SLMs: Developing small language models (SLMs) for speech, Occupational Therapy (OT), and special education, focusing on specific needs.

VI. Bias Mitigation & Data Integrity

• Bias Identification: Implementing bias identification using fairness metrics to detect and mitigate bias in our deployed models.
• Bias Mitigation Techniques: Utilizing bias mitigation techniques, including data pre-processing and algorithm modification, to address bias in classification.
• Data Diversity & Monitoring: Emphasizing diverse data and continuous monitoring to ensure representative data and promote fairness.

VII. Prompting & Refinement

• Structured Prompting: Using structured prompting with clear, specific prompts to guide AI responses for targeted information and support.
• Context & Persona: Leveraging context and persona in prompts, tailoring AI interactions for specific client needs and scenarios.
• Output Refinement: Refining AI outputs through exemplars and tone prompting to ensure accuracy, relevance, and ethical considerations.

VIII. Conversation Management

• Organized Conversations: Organizing and labeling conversations for easy access and reference, improving efficiency and knowledge management.

IX. Data Architecture

• Data Storage: PostgreSQL database with normalized schema for operational data, coupled with robust metadata management for discoverability and governance.
• Data Warehouse: A dedicated data warehouse stores aggregated and transformed data, enabling analytical reporting, performance monitoring, and business intelligence insights.
• Knowledge Graph: A property graph database (e.g., Neo4j) represents relationships between neurodiversity concepts, resources, experts, and user profiles, facilitating knowledge discovery.
• Vector Database: Prach utilizes Pinecone, a scalable vector database, with OpenAI embeddings and cosine similarity to power semantic search and retrieval for RAG.
• Data Pipeline: The data pipeline includes validation against schemas, user-generated content moderation, data extraction from diverse sources, and transformation using Python tools.
• RAG: The Retrieval Augmented Generation process involves generating and storing embeddings, retrieving relevant chunks from the vector database and knowledge graph, and augmenting LLM prompts.
• Data Access: Prach employs Role-Based Access Control (RBAC), data federation for external sources, OAuth 2.0 authentication, and manages separate RAG instances for data privacy and relevance.