sanjay Azhagan
AI & Full-Stack Engineer shipping fast, scalable MVPs.
New to Contra
sanjay is ready for their next project!
Built a legal-document auditing system that automates PII redaction and contract risk assessment. Implemented LLM workflows using LangGraph and Map-Reduce techniques to scan corporate agreements for material risks like uncapped liabilities. Backend built with Python and FastAPI, frontend with Node.js/React, designed for accuracy and reliability on sensitive legal data. Solves a real problem: manual contract review is slow and error-prone, this automates the first-pass risk scan so lawyers focus only on flagged issues.
Youtube demo :https://www.youtube.com/watch?v=4te5t5ihVNs
live link:https://sanjayazhagan.tech/veri-law/
1
22
Full-stack portfolio website built with Next.js 16, React 19, TypeScript, and Tailwind CSS. Implemented secure authentication with NextAuth and bcrypt. Prisma ORM manages SQLite database with user profiles and project management. Key features: PDF resume generation, Cal.com (http://Cal.com) booking integration, PostHog analytics, Mermaid diagram support, and Markdown content rendering. Advanced animations with Framer Motion. Demonstrates expertise in modern web development: responsive design, backend API development, database design, and production deployment.
0
15
Problem: LLM fine-tuning often trades accuracy for format consistency. Models trained on reasoning tasks frequently output inconsistent formats, requiring expensive post-processing pipelines and manual validation before production use.
Solution: Built an autonomous math-reasoning system using GRPO (Group Relative Policy Optimization) that fine-tunes Qwen2.5-0.5B-Instruct with a verifier-based reward system. The model learns to solve grade-school math problems while maintaining 100% structured output format.
Results:
Format compliance: 100% (structured tags like <answer>X</answer>)
Math accuracy: 48% (significant improvement on GSM8K)
Inference cost: ~0.15¢/request (lightweight 0.5B model)
Why it matters: Perfect for education platforms, AI tutoring APIs, or any system requiring both accurate reasoning AND reliable output structure.
Tech: Qwen2.5-0.5B, GRPO, verifiers package, Hugging Face Transformers, Kaggle T4 GPUs
0
13
Problem: Traditional study groups suffer from the "freeloader problem"—students benefit from shared notes but rarely contribute quality content. Plus, generic AI assistants provide irrelevant answers across different subjects.
Solution: StudySphere is a pool-based, token-incentivized RAG platform where students earn tokens by uploading high-quality notes and spend tokens asking AI questions—creating a self-sustaining peer-learning ecosystem.
How It Works:
Isolated Context Pools: Each subject has its own RAG database. When you ask a question, the AI only draws from that pool's notes—no irrelevant hallucinations
Token Economy: Upload notes → earn tokens → ask AI questions → spend tokens. Higher quality notes earn more tokens
AI Auto-Grading: Every PDF is automatically scored by Gemini for quality, depth, and clarity before being added to the pool
Built For: Universities, study groups, tutoring centers, or any peer-learning environment where context accuracy matters (medical, law, competitive exams)
Tech Stack: Node.js, Express, PostgreSQL, Gemini API, LangChain, Pinecone (vector DB with namespaces for isolation), JWT auth
0
7