Digital krishna

• 

  Saathi on mobile—multilingual, private and practical guidance designed for everyday concerns.

• 

  Digital Krishna turns timeless wisdom into calm, culturally grounded next steps for modern life.

• 

  Explore the Bhagavad Gita, Ramcharitmanas, Upanishads and Krishna stories in one scripture library.

• 

  Saathi GPT uses a fine-tuned Qwen model to provide structured, multilingual and responsible guidance.

• 

  A responsive mobile experience that keeps guidance, scripture and reflection accessible anywhere.

• 

  Guidance in action—relevant wisdom, source context and practical steps for the user’s real concern.

• 

  Read verified scripture on mobile with clear Hindi and English meanings and accessible navigation.

• 

  Story-based reflection connects personal challenges with memorable lessons from Krishna’s life.

• 

  A unified scripture library for focused reading, reflection and culturally grounded learning.

• 

  Visual storytelling makes Krishna’s teachings engaging, memorable and easier to apply in daily life.

Inspiration

Young people today face stress, academic pressure, loneliness, anger, uncertainty, family difficulties, career confusion, and questions about purpose. Although general-purpose AI systems can respond to these problems, their answers are often generic, culturally disconnected, inconsistent, or limited to motivational advice.

Digital Krishna was inspired by the relationship between Krishna and Arjuna in the Bhagavad Gita. Krishna did not simply give Arjuna an answer; he helped him understand his situation, examine his responsibilities, regain clarity, and take thoughtful action.

We wanted to explore whether a language model could be specially trained to provide similarly structured, culturally familiar, multilingual, and practical guidance for modern-life challenges.

Digital Krishna is not an AI claiming to be God, a therapist, or a replacement for professional care. It is a domain-specific AI system inspired by Krishna’s teachings and designed to make timeless wisdom understandable and actionable for modern users.

What it does

Digital Krishna is a supervised fine-tuned language model built on Qwen 4B.

Instead of relying only on prompts around a general-purpose model, we created a custom instruction dataset that teaches the model how to respond to different categories of user problems.

The model is designed to support conversations involving:

• Study pressure and lack of discipline
• Career and decision-making difficulties
• Anger, grief, loneliness, and emotional confusion
• Family and relationship challenges
• Self-confidence and comparison
• Spiritual questions and questions about life purpose
• Ambiguous or incomplete user messages
• Multi-turn conversations and user corrections
• English, Hindi, and natural Hinglish interactions

Digital Krishna follows a structured guidance approach:

1. Understand the user’s situation
2. Respond with empathy without unnecessary judgement
3. Clarify missing information when required
4. Connect the situation with an appropriate Krishna-inspired teaching
5. Explain the teaching in simple modern language
6. Provide practical actions the user can take
7. Ask a useful follow-up question when appropriate

The website provides the user interface, while the supervised fine-tuned model is the core technology behind the experience.

Why supervised fine-tuning?

A normal prompt can temporarily tell a model how to answer, but it may not provide consistent behaviour across thousands of different situations.

Supervised fine-tuning allows Digital Krishna to learn:

• A consistent response structure
• Emotionally appropriate tone
• Culturally grounded guidance
• Multilingual conversation patterns
• When to provide a concise or detailed response
• When to ask for clarification
• When not to force spiritual references
• How to convert philosophical ideas into practical next steps
• How to handle sensitive situations responsibly

This makes Digital Krishna different from a standard chatbot with a spiritual system prompt.

How we built it

We started with the Qwen 4B language model and prepared a custom supervised fine-tuning dataset containing varied instruction-response conversations.

The dataset includes:

• Basic conversations and greetings
• Single-turn and multi-turn discussions
• Ambiguous follow-up questions
• Emotional and practical life problems
• English, Hindi, and Hinglish examples
• Short, medium, and detailed response styles
• Multiple valid answers for similar problems
• Safety-focused examples
• Situations where clarification is required
• Examples where spiritual guidance is relevant and where it is unnecessary

The training data was reviewed to reduce repeated phrases, rigid templates, unsupported claims, and overly generic responses.

Our system architecture includes:

User
  ↓
Digital Krishna Web Interface
  ↓
Input and Safety Processing
  ↓
Supervised Fine-Tuned Qwen 4B Model
  ↓
Verified Knowledge and Guidance Layer
  ↓
Personalized Response
  ↓
Text, Voice, Stories, and Practical Actions

Built With

• amazon-web-services
• database
• face
• fine-tuning
• hugging
• lora
• natural
• next.js
• peft
• python
• pytorch
• qlora
• qwen
• rag
• react
• supabase
• supervised
• text-to-speech
• transformers
• typescript
• vector