
SOM.ai
Vedic intelligence engine grounded in the Bhagavad Gita. The AI system that delivers philosophical clarity for complex decisions. No generic advice, no filler just timeless wisdom.
Timeline
7 Weeks
Role
Full Stack AI Engineer
Team
Solo
Status
CompletedTechnology Stack
Key Challenges
- RAG Pipeline over a Fixed Philosophical Canon
- Multi-Agent Reasoning Chain Architecture
- pgvector-Backed Semantic Search
- Context-Aware Verse Retrieval
- Structured Output (Meaning → Apply → Reflect)
- Conversation Memory Across Sessions
- Safety Moderation Layer for Sensitive Queries
Key Learnings
- Retrieval-Augmented Generation (RAG) Design
- pgvector Embedding & Similarity Search
- Multi-Agent Orchestration Patterns
- LLM Prompt Engineering for Philosophical Reasoning
- Next.js 14 App Router Architecture
- Grounding LLMs in a Fixed Knowledge Canon
- Output Structuring for Actionable AI Responses
Summary
"Think Clearer. Decide Better." SOM.ai is a Vedic intelligence engine grounded in the Bhagavad Gita , a philosophical reasoning system built for people who think for a living. No surface-level advice. No generic platitudes. Just timeless clarity. With 700 verses indexed and instantly searchable, an 8-agent reasoning chain, and a RAG pipeline backed by pgvector, SOM.ai transforms your real-world struggles into grounded, actionable wisdom in under 5 seconds.
Features
- Vedic Intelligence Core — Every answer is matched to a specific Bhagavad Gita verse, not scraped from random web pages
- Structured Output — Responses follow a three-part format: Meaning Synthesis → Apply Today → Reflect
- Multi-Agent Reasoning — 8 specialized agents collaborate on context retrieval, synthesis, and moderation
- RAG + Vector Search — pgvector-backed semantic search across all 700 Gita verses for precise retrieval
- Conversation Memory — Remembers your full journey across sessions, not just the last few messages
- Safety Moderation — Active moderation layer ensures responsible, grounded output for sensitive queries
- Instant Recall — Sub-5-second response time from query to philosophical insight
Architecture with Real-Life Use Case
The Vedic Intelligence Pipeline
When a user says: "My startup is pivoting and I feel like a failure."
- Query Intake & Sanitization (instant) → Input is parsed and moderated for context
- Semantic Verse Retrieval (< 1 sec) → pgvector search finds the most relevant Gita verse(s)
- Multi-Agent Reasoning Chain (2–3 sec) → 8 agents collaborate on meaning, application, and reflection
- Structured Synthesis (instant) → Output is formatted as Meaning → Apply → Reflect
- Moderation Pass (instant) → Safety layer validates response before delivery
- Session Memory Update (instant) → Conversation context is persisted for continuity
Total: < 5 seconds vs. years of trial-and-error, therapy queues, or hollow internet advice
Real-World Outcome
Generic AI Approach:
"Take a deep breath." "Have you tried journaling?" "Consider talking to someone."
Surface-level. Forgettable. Treats the symptom, not the cause.
SOM.ai Approach:
User asks: "My startup is pivoting and I feel like a failure."
Results in < 5 seconds:
- Verse Retrieved: Bhagavad Gita 2.47 — "You have a right to perform your prescribed duty, but you are not entitled to the fruits of action."
- Meaning Synthesis: Anxiety stems from obsessing over outcomes outside your control. Reclaim agency by focusing strictly on execution.
- Apply Today: Define success today entirely by the effort expended on your deepest work block — disregard external validation.
- Reflect: "What outcome are you mentally attaching to right now that is causing friction?"
✅ Result: Root-cause clarity grounded in 5,000 years of philosophical truth.
Tech Stack
- Frontend: Next.js 14 (App Router), TypeScript, Tailwind CSS
- Backend: FastAPI, Python, SQLAlchemy (async)
- AI / RAG: Groq API (LLaMA), pgvector, Retrieval-Augmented Generation
- Database: PostgreSQL with pgvector extension
- Architecture: Multi-agent orchestration, context-aware synthesis, moderation layer
Key Technical Achievements
- Fixed Philosophical Canon: Unlike general-purpose LLMs, SOM.ai is grounded exclusively in the 700 verses of the Bhagavad Gita — eliminating hallucinations and ensuring every response has a traceable source
- 8-Agent Reasoning Engine: Specialized agents handle retrieval, synthesis, application framing, reflection prompting, memory management, moderation, and more
- pgvector Semantic Search: Verse embeddings are stored in PostgreSQL via pgvector, enabling meaning-level similarity search far beyond keyword matching
- Structured Output Format: Every response is intentionally shaped into three actionable layers — Meaning, Apply, and Reflect — making wisdom immediately usable
- Conversation Memory: Full session history is retained, enabling SOM.ai to track your mental and philosophical journey over time
Architecture Highlights
Vedic RAG Pipeline:
User Query → Sanitization → Embedding Generation →
pgvector Semantic Search → Verse Retrieval →
Multi-Agent Reasoning Chain → Structured Synthesis →
Moderation Pass → Final Response
Architecture Comparison
-
Where answers come from
- Typical AI: Scraped from random web pages
- SOM.ai: Matched to a specific Gita verse
-
What you actually get
- Typical AI: A wall of text
- SOM.ai: Meaning → Apply → Reflect
-
Conversation memory
- Typical AI: Forgets after a few messages
- SOM.ai: Remembers your full journey
-
Depth of thinking
- Typical AI: One-shot answer
- SOM.ai: Multi-agent reasoning chain
Real-World Impact Example
Founders, Engineers & Therapists
The Problem with Generic AI:
- Produces engagement-optimized answers, not truth-optimized ones
- No philosophical grounding — answers shift based on phrasing
- Surface-level coping advice that treats symptoms, not root causes
- Forgets context between sessions; no continuity of thought
With SOM.ai:
- Every answer is traceable to a specific Gita verse with a 5,000-year track record
- Multi-agent reasoning pinpoints the root cause of your mental friction
- Structured output gives you something concrete to act on today
- Session memory means SOM.ai grows more useful the more you use it
Built for people who think for a living : the founders making high-stakes pivots, the engineers debugging their decision-making, the therapists seeking evidence-backed philosophical frameworks for their clients.
Development Journey
Technical Challenges Overcome
-
Grounding an LLM in a Fixed Canon
- Problem: General LLMs hallucinate and drift from authoritative sources
- Solution: RAG pipeline strictly limits retrieval to the 700 indexed Gita verses via pgvector similarity search
-
Multi-Agent Coordination at Speed
- Problem: Chaining 8 agents risks compounding latency and coherence loss
- Solution: Async FastAPI architecture with parallel agent execution where possible, keeping responses under 5 seconds
-
Meaning-Level Verse Retrieval
- Problem: Keyword search misses semantically relevant verses
- Solution: Sentence-level embeddings stored in pgvector enable meaning-aware retrieval even for abstract emotional queries
-
Structuring Philosophical Output for Action
- Problem: Raw verse translations are cryptic and hard to apply to modern problems
- Solution: Dedicated synthesis and application agents transform ancient text into a three-part actionable framework
-
Conversation Memory Without Context Bloat
- Problem: Full history injection bloats the context window and slows reasoning
- Solution: Memory agent maintains a compressed session summary that preserves continuity without sacrificing speed
Performance Benchmarks
-
Average Response Time
- Target: < 10 sec
- Achieved: < 5 sec
-
Verses Indexed
- Target: 700
- Achieved: 700 (complete Bhagavad Gita)
-
Moderation Coverage
- Target: Key sensitive categories
- Achieved: Active moderation on every response
Resources & Links
- Live Platform: som.ai.thevanshgarg.com
- GitHub: github.com/vanshxdevs
- Free Access: Start free , no credit card required, takes 30 seconds
For Employers & Collaborators
What This Project Demonstrates:
- RAG Systems Engineering: Production-grade retrieval pipeline grounded in a fixed philosophical canon with pgvector
- Multi-Agent Architecture: 8-agent orchestration system handling retrieval, synthesis, application, memory, and moderation
- LLM Prompt Engineering: Structured output design translating ancient text into modern, actionable three-part responses
- Semantic Search: pgvector-backed embedding search enabling meaning-level verse retrieval from a 700-verse corpus
- Product Philosophy: Opinionated AI product design and grounding intelligence in truth, not engagement optimization
- Full-Stack Ownership: End-to-end architecture from RAG pipeline to Next.js 14 frontend
Contact: Available for technical interviews or architecture deep-dives
Project Status
Current: Live at som.ai.thevanshgarg.com
Corpus: 700 Gita verses indexed • 8 reasoning agents • < 5s response time
Next Milestones:
- Expanded verse corpus with Upanishads and Stoic philosophy
- Personalized insight tracking and journaling layer
- Team / shared session features for coaching use cases
- API access for therapists and mental health practitioners