AI Feynman Kannada Tutor — Vavi Labs

Building a Kannada Physics Tutor LLM with Feynman-Style Explanations

Multi-stage fine-tuning pipeline to create a reasoning-first physics tutor in Kannada, combining domain-specific SFT and RAG for intuitive, grounded explanations.

Dataset on HuggingFace ↗ Kannada General SFT ↗ Kannada Physics SFT ↗

Why This Project?

Most LLMs struggle with conceptual physics explanations in regional languages. This project demonstrates three things at once:

Low-Resource Language

Teaching a model to think in Kannada, not just translate from English.

Feynman-Style Teaching

Intuitive, step-by-step reasoning with analogies — focusing on why, not just formulas.

Multi-Stage Pipeline

Language fluency → Domain reasoning → Factual grounding with RAG.

This case study demonstrates:

End-to-end fine-tuning for low-resource languages + specialized domains
How multi-stage SFT compounds improvements (language → domain → grounding)
Complete evaluation pipeline: LLM-as-judge, quantitative metrics, qualitative analysis

System Overview

Goal: Given a physics question in Kannada (e.g., "How do we derive Laplace's equation for a membrane?"), generate:

A step-by-step reasoning trace that builds intuition.
A Feynman-style explanation in natural, fluent Kannada.
Grounded answers using retrieved physics context (RAG).

Four-Model Progression

The training pipeline uses a staged approach where each model builds on the previous one:

1. Base Model

Gemma 3 1B — General-purpose LLM baseline (no specialization)

2. Kannada SFT

Language Foundation — Fine-tuned on general Kannada text for fluency

3. Physics SFT

Domain Reasoning — Physics concepts + Feynman-style explanations in Kannada

4. Physics + RAG

Factual Grounding — Retrieval-augmented generation with physics knowledge base

Why Multi-Stage Fine-Tuning?

Stage 1 (Kannada SFT): Teaches linguistic fluency independent of domain.

Stage 2 (Physics SFT): Builds conceptual reasoning and Feynman-style intuition on top of that fluency.

Stage 3 (RAG): Adds factual precision by retrieving relevant physics context before answering.

LLM-as-Judge Evaluation

Each model's output is evaluated using an LLM-as-judge with a structured grading prompt. Scores are on a 0–5 scale across four dimensions:

Physics Correctness

Conceptual accuracy and mathematical precision

Reasoning Quality

Step-by-step clarity and logical flow

Kannada Fluency

Natural language quality and proper grammar

Contextual Relevance

Answers the specific question with appropriate detail

Quantitative Evaluation

All four models were evaluated on a held-out set of physics questions, scoring each output with an LLM-as-judge (0–5 scale).

Score Distribution by Model

Progressive right-shift in score distributions demonstrates that each stage adds measurable value: fluency → reasoning → grounding.

Boxplot Comparison

Median score increases dramatically: Base (≈0) → Kannada SFT (0.5) → Physics SFT (2.3) → RAG (3.7). The IQR narrows for final models, showing more consistent performance.

Key Takeaway: Multi-stage fine-tuning doesn't just improve average performance — it systematically reduces failure modes and increases reliability.

Hero Examples

Compare the output quality of all four models on held-out physics questions. Scores are on the 0–5 LLM-as-judge scale.

Detailed Training Specifications

Training Pipeline Overview

Base model → Kannada SFT → Physics SFT → RAG integration. Each stage compounds improvements from the previous.

Data Preparation

Base ModelGemma 3 1B (google/gemma-3-1b-it)

Kannada SFT DataGeneral Kannada text corpus (news, essays, conversations)

Physics SFT DataCustom physics Q&A with Feynman-style reasoning traces in Kannada

RAG Knowledge BasePhysics textbook chapters in Kannada (vector store)

Training Configuration

Hardware: 1× A40 (48 GB VRAM) · PyTorch with HuggingFace Transformers

Stage 1: General Kannada SFT

Training Steps80

Epochs1

Batch Size8

Max Learning Rate5e-5

Initial Training Loss4.709

Final Training Loss2.898

Final Eval Loss2.912

Eval IntervalEvery 5 steps

Stage 2: Physics Kannada SFT

Training Steps90

Epochs1

Batch Size8

Max Learning Rate5e-5

Initial Training Loss3.248

Final Training Loss2.448

Final Eval Loss2.474

Eval IntervalEvery 10 steps

Training Progression:

Stage 1 reduced eval loss from 4.71 → 2.91 (38% reduction)
Stage 2 further reduced eval loss from 3.25 → 2.47 (24% reduction)
Healthy train/eval gap indicates good generalization without overfitting

Key Learnings

1

Low-Resource Language Fine-Tuning

Separating language fluency (Kannada SFT) from domain knowledge (Physics SFT) allows the model to learn each skill independently before combining them. This staged approach is more effective than trying to learn both simultaneously.
2

Multi-Stage SFT Compounds Benefits

Each stage adds measurable value: language quality → reasoning patterns → factual accuracy. The boxplots clearly show each stage shifts the performance distribution upward.
3

RAG as a Grounding Layer

Fine-tuned models can reason well but may still hallucinate details. Adding RAG dramatically improves factual accuracy without additional training.
4

LLM-as-Judge for Evaluation

Building a complete evaluation pipeline with LLM-as-judge, quantitative metrics, and qualitative analysis provides a full picture of model capability.

From Demo to Product: AI-Powered Physics Education for Rural India

This trained model demonstrates the foundation for a scalable educational platform designed for government schools and rural communities across India where students lack access to quality STEM education in their native language.

Instead of replacing teachers, the platform acts as an always-available AI tutor that supplements classroom learning and helps students build deep conceptual understanding in their mother tongue.

The Challenge in Rural India

Language Barrier

In rural Karnataka, 85% of students study in Kannada-medium schools, but most quality educational resources are only available in English.

Teacher Shortage

Government schools face chronic shortages of qualified physics teachers, leaving little time for conceptual depth or individual attention.

Rote Learning Culture

Traditional teaching emphasizes memorizing formulas without understanding — exactly what Feynman-style teaching addresses.

Cost of Tutoring

Private physics tutoring costs ₹3,000–5,000/month — unaffordable for families earning ₹15,000/month.

Multi-Agent AI Tutor Architecture

1

Concept Explainer Agent — Feynman-Style Intuitive Teaching

This is where our Physics+RAG model powers the system. Generates Feynman-style explanations with analogies, step-by-step reasoning, and visual intuition in Kannada.
2

Practice Problem Generator — Adaptive Problem Sets

Generates chapter-aligned practice problems with difficulty progression and detailed worked solutions in Kannada.
3

Learning Path Advisor — Personalized Curriculum

Tracks student progress, identifies conceptual gaps, and recommends targeted review topics before advancing.
4

Exam Preparation Coach — Board Exam Focused Support

Targeted support for Karnataka SSLC and PUC board exams: previous year question analysis, mock tests, and answer writing practice.

Revenue Model: Freemium — basic concept explanations free for all students. Premium features at ₹99/month ($1.20). Government partnerships and NGO sponsorships subsidize access below poverty line. Same multi-stage fine-tuning approach can be replicated for Hindi, Tamil, Telugu, and other regional languages.