III Year — Undergrad
VARUN KUMAR
Artificial Intelligence · Deep Learning · Agentic Systems
Building intelligent systems that think, learn, and adapt.
From multi-agent architectures to deep learning pipelines —
I engineer the future of AI, one model at a time.
⚡ Languages & Tools
Python
Java
SQL
MongoDB
PostgreSQL
MySQL
Redis
Qdrant
Docker
Git
GitHUB
Competitive Programming
Neo4j
LangChain
🧠 AI & Machine Learning
Machine Learning
Deep Learning
Generative-AI
RAG
Agentic RAG
Graph-RAG
Fine-Tuning
Agentic AI
LLM
Multi-Agents
Google Multi-Agent Modules
🏆
Synaptix AI Hackathon — IIT Madras
Finalist
💻
Competitive Programming — IIT Madras
Finalist
🚀
TATA Imagination Challenge
Semi-Finalist
Built an enterprise-grade Retrieval-Augmented Generation (RAG) pipeline designed for
high-performance semantic search and autonomous document ingestion.
Engineered a sophisticated multi-tier caching architecture
(L1/L2/L3) to drastically reduce latency and improve query response times
Built automated pipelines for web scraping, PDF parsing, and
chunking, seamlessly integrated with a custom Composite File System (CFS)
Implemented Rotary Position Embeddings (RoPE) and integrated
Qdrant Vector Database for highly scalable and accurate semantic search
Designed robust data integrity mechanisms utilizing
Write-Ahead Logging (WAL) and adaptive task scheduling
Python
Qdrant Vector DB
NLP
RoPE Embeddings
PyTest
Designed and simulated an AI-powered Hospital IVR system replacing legacy DTMF navigation with a
hybrid ILP-AI architecture for automated symptom triage and emergency ambulance dispatch.
Architected a hybrid ILP + SLM pipeline using Gemini-2.5-flash
for
clinical reasoning and SCIP for optimal ambulance-to-patient assignment
Integrated Chirp embedding model for speech-to-text conversion
and
intent detection
Designed an NLU pipeline with speech-to-text, intent
detection, distress classification, and clinical NER for multi-language scenarios
Simulated a microservices backend (NestJS + Python) with
PostgreSQL and Redis, handling 1,000+ concurrent call scenarios at 99.9% uptime
Gemini-2.5-flash
NER Models
Chirp Embedding Model
NestJS
Python
FastAPI
PostgreSQL
Redis
HL7/FHIR
Built an AI-powered skill intelligence platform to enrich labor market data using knowledge
graphs and LLMs.
Constructed a Neo4j knowledge graph with 2,800+ skill nodes
and 768-dim embeddings capturing skill taxonomy hierarchies
Developed a job description enhancement pipeline using Gemini
1.5 Pro with few-shot prompting and chain-of-thought reasoning
Trained a Graph Neural Network (GNN) with Node2Vec embeddings
for skill classification and relationship inference
Neo4j
Gemini 1.5 Pro
GNN
spaCy
Node2Vec
Worked on building a scalable AI-powered document query system for enterprise users, enhancing
document search and retrieval using NLP and vector search technologies.
Developed a full-stack RAG system using FastAPI, Streamlit,
MongoDB, Qdrant, and Hugging Face Transformers
Implemented real-time semantic search across multiple PDFs
with support for voice-based queries
Improved document search speed by 80% with enterprise-grade
concurrent request handling
FastAPI
Streamlit
MongoDB
Qdrant
Hugging Face
Contributed to the end-to-end development of a Restaurant Management System tailored for a local
business client, focusing on order processing, menu management, and billing features.
Developed the backend using Java and MySQL, and built user
interfaces with Java Swing
Designed modules for order tracking, table reservation, and
inventory management
Delivered the solution with 100% on-time delivery and zero
post-deployment bugs
Java
MySQL
Java Swing
SDLC
✦
VOID LOG
Thoughts, experiments, and dispatches from the infinite void.
Enter the Void Log →
