AWS Case Study

Teertham: Vedic Astrology Chatbot on Bedrock

Case Description Teertham, India’s trusted spiritual services platform, partnered with a21.ai to build “AI Pandit” — a RAG-based conversational astrology chatbot on Amazon Bedrock. The system grounds responses in classical Vedic texts (BPHS, Saravali, Phaladeepika), performs Kundali-based puja recommendations via tool-calling, and includes a multi-

Case Study
PLACEHOLDER DOCUMENT

Teertham: Vedic Astrology Chatbot on Bedrock Case Description: Teertham, India's trusted spiritual services platform, partnered with a21.ai to build "AI Pandit" — a RAG-based conversational astrology chatbot on Amazon Bedrock. The system grounds responses in classical Vedic texts (BPHS, Saravali, Phaladeepika), performs Kundali-based puja recommendations via tool-calling, and includes a multi-layered trust & safety framework for emotionally vulnerable users — converting spiritual curiosity into puja bookings. Problem Statement: Teertham runs a spiritual services marketplace where families book verified Pandits for 100+ ritual types (Griha Pravesh, Mundan, Satyanarayan Katha, Grah Dosh Nivaran, etc.). The platform faced several growth bottlenecks: - Pre-booking drop-off: users landed on puja listings but couldn't determine which of 100+ pujas was appropriate for their situation (new job, health concern, Grah Dosh). No guidance available outside WhatsApp support hours. - Manual support overhead: the only support channel was WhatsApp, creating manual load and limiting async availability. Target: 60-70% deflection via AI. - Terminology barrier: first-generation diaspora users and urban millennials lack ritual vocabulary to search/navigate puja types effectively. - Muhurat friction: auspicious timing questions required a human Pandit call before booking, adding days of latency. - Kundali interpretation gap: users had birth charts generated but no guidance on what remedies their chart indicated. - Emotional safety risk: users in grief or distress (bereavement, fear, anxiety) needed structured, safe pathways — not astrological predictions. Proposed Solution & Architecture: a21.ai built "AI Pandit" — a RAG-based conversational chatbot deployed on AWS in ap-south-1 (Mumbai). AI/ML Tier — Amazon Bedrock: - Amazon Bedrock: Anthropic Claude Sonnet 4.6 for primary reasoning, tool-calling, and response synthesis. Multi-turn conversation with 6-layer system prompt architecture (Persona → Safety → Knowledge Grounding → Intent Classification → Tool Usage → Response Format). - Amazon Bedrock Embeddings: Titan Embeddings v2 for knowledge chunk vectorization. - Amazon Bedrock Guardrails: content filters for harmful spiritual predictions, PII protection for Kundali data (birth date/time/place classified as sensitive personal data under India's DPDP Act 2023), denied topics (political, sectarian, inter-religious comparisons). RAG Knowledge Base — Vedic Astrology Corpus: Classical texts ingested, chunked (1,500 chars / 200 overlap), and vectorized: - Brihat Parashara Hora Shastra (BPHS): foundational Vedic astrology (Kundli, Dasha, Yogas) - Teertham puja catalog: 100+ puja descriptions, Samagri lists, pricing, durations - Generated Q&A pairs: Hinglish training data for culturally appropriate responses Vector store: Amazon RDS PostgreSQL 15+ with pgvector extension. Tool-Calling (Agentic Capabilities): Bedrock Claude invoked with tool-use for: - get_detailed_kundli_insights: fetches user's birth chart data (Dashas, house lords, planetary positions) from UserKundlis table, pruned to 19 keys to stay within context window - get_muhurat_data: queries Panchang API for auspicious dates relevant to user's query - get_puja_details: retrieves specific puja info (Samagri list, pricing, duration, booking deeplink) - create_booking_deeplink: generates direct booking URL for recommended puja Intent Classification System: Every query classified before response generation: - Puja recommendation / Ritual guidance / Kundali interpretation / Muhurat query / FAQ (pricing, duration) / Emotional distress (Category A — overrides all other intents) Trust & Safety Framework (6 Layers): 1. Distress keyword classifier: runs BEFORE main LLM pipeline. If triggered, response template overrides KB content. Escalates to human Acharya via WhatsApp. 2. Mental health boundary: chatbot does not reinforce delusional/dissociative beliefs. 3. Fear-based query handling: deterministic negative predictions strictly forbidden. All Dosha/astrology responses include: "This is general guidance. For personalised advice, consult one of our verified Acharyas." 4. Hallucination prevention: confidence-threshold fallback to human Acharya. Wrong mantras, wrong deity associations, fabricated Muhurats, incorrect pricing all classified as critical hallucination risks. 5. Content governance: all mantra content reviewed by Sanskrit scholar. No political/sectarian content. User feedback widget with 48-hr review SLA. 6. Guardrails: Bedrock Guardrails enforce PII redaction, denied topics, and content filters at inference layer. Infrastructure: - Amazon ECS (Fargate): Node.js 18+ admin service (Express + Sequelize) + FastAPI AI service (Python, LangChain/LangGraph agent orchestration) - Amazon RDS PostgreSQL 15: application DB + pgvector knowledge chunks + UserKundlis - Amazon S3: document storage (puja images, Kundali PDFs, Vedic text source PDFs), static assets via CloudFront - Amazon CloudWatch: observability (token consumption, tool-call frequency, distress detection triggers, conversation metrics) - Amazon SNS: Slack/WhatsApp notifications for distress escalation and human handoff Project Outcomes and Success Metrics: AI Pandit was designed primarily as a revenue-growth and customer-engagement platform. By providing always-on spiritual guidance, personalized puja recommendations, and direct booking pathways, the solution is projected to improve booking conversion from 2% to 3%. - Increase Conversion from 2% to 3% - Estimated ROI: $216K / $49.1K = 4.4x (~440%) - Estimated payback period: ~7 months Teertham PRD-Defined Targets: - Support deflection: 60-70% of WhatsApp queries handled by AI Pandit - Booking conversion lift: puja recommendation to booking deeplink conversion measured - Hallucination rate: <1% (wrong mantra, wrong deity, wrong Muhurat, wrong pricing) - Distress detection accuracy: 100% recall on Category A keywords (false negatives = safety failure) - User satisfaction: maintain 98%+ platform satisfaction score post-chatbot launch - Availability: 24/7 (vs. WhatsApp human support limited to business hours)