index github.com/any/repository
- 01Parse — tree-sitter AST, 36 languages
- 02Embed — Jina AI semantic vectors
- 03Retrieve — grounded context, not guesses
- 04Reason — Claude on Bedrock, 6-round loop
- 05Stream — answers over SSE
I'm Devanshu Verma — a Data Science & AI undergrad at IIIT Naya Raipur who ships agentic, RAG-powered products end-to-end. My flagship, Codrix.ai, indexes any GitHub repo and answers questions about it conversationally.
index github.com/any/repository
About
Retrieval that actually retrieves. Agent loops that know when to stop. Backends that degrade gracefully — exponential backoff, automatic fallbacks, real-time streaming. That's where I spend my time.
Pre-final year, Data Science & Artificial Intelligence at IIIT Naya Raipur — turning coursework in ML, systems, and algorithms into shipped software like Codrix.ai.
RAG pipelines, LLM orchestration, tool-use agents — and evaluation that keeps them honest.
Go and Node services on AWS — streaming APIs, retry policies, fallbacks designed first.
Data Science & AI, expected May 2028 · CGPA 8.67/10. Coursework applied directly to real builds.
Stack
A stack assembled project by project — weighted toward agentic AI and retrieval, with deliberate depth in the languages and infrastructure underneath.
Daily driversPython · C++
Vector memoryChromaDB · Pinecone · FAISS
Ships withDocker · Git
FocusAgentic AI & RAG
Work
My flagship project: a code intelligence platform that indexes any public GitHub repository, answers questions about it conversationally, and draws its diagrams on demand.
A custom code intelligence engine using tree-sitter AST parsing across 36 programming languages, paired with Jina AI vector embeddings for semantic code search over any repository.
Large repos are opaque — onboarding means scrolling files, guessing architecture, and pinging teammates. Codrix makes a repo explain itself.
Exponential backoff on Bedrock calls, automatic Groq LLaMA fallback, and Server-Sent Events so long agent runs stay transparent — never a frozen spinner.
A Prompt-to-Diagram engine renders Architecture, UML Sequence, and ER views from the actual codebase — then refines them in plain English.
Journey
Amazon's AI for Bharat Hackathon — a nationwide competition for AI-driven solutions for India, judged on real working prototypes.
Expected May 2028 · CGPA 8.67/10. Foundations in ML, algorithms, and systems — applied directly to personal builds like Codrix.ai.
Advanced to the prototype submission phase of Amazon's AI for Bharat Hackathon — presented Codrix.ai end-to-end, built on AWS Bedrock, EC2, and the Converse API.
Real-time streaming, multi-round agent reasoning, and multi-language code intelligence — deployed, demoed, and still evolving.
I'm actively open to work and looking for internships where I can build real AI systems end-to-end. Openings, research problems, hackathon teams, agentic products — I reply fast.
devanshuverma372@gmail.com · github.com/DeV-433
Open to internships
AI Engineer · Agentic systems & RAG
Pre-final year B.Tech student in Data Science & AI at IIIT Naya Raipur. I build agentic, RAG-powered products end-to-end — most recently Codrix.ai, a code intelligence platform built for Amazon's AI for Bharat Hackathon.
Code intelligence platform · Amazon AI for Bharat Hackathon
Indexes any public GitHub repository and answers questions about it conversationally.
Go · React 18 · Node.js · AWS Bedrock · tree-sitter · RAG · Docker
Amazon Web Services · Nationwide
Advanced to the prototype submission phase; built and presented Codrix.ai on AWS Bedrock, EC2, and the Bedrock Converse API — a production-grade RAG pipeline with real-time streaming.
B.Tech, Data Science & Artificial Intelligence · CGPA 8.67/10
Coursework: Data Structures & Algorithms, Machine Learning, Foundations of AI, Reinforcement Learning, Statistical Data Analysis, Data Mining, DBMS, Operating Systems, OOP.
Have an internship, research problem, or team that needs an AI builder? I reply fast.
devanshuverma372@gmail.comWant the full story and how it all works?