// portfolio · build 2026.09
I build AI agents, the evaluation pipelines that keep them honest, and retrieval systems that hold up on real data.
from ratul import nownow(){
'study': 'B.Tech CSE · VIT-AP',
'focus': [
'Backend for AI',
'Evaluation & Retrieval Systems',
'Multi-Agent Workflows',
],
'cgpa': 8.62,
'grad': 2027,
}I'm a final-year Computer Science student at Vellore Institute of Technology, Andhra Pradesh, specializing in Data Analytics, and an AI Engineer intern at TailorTalk. Most of my time goes into generative AI systems that have to work on real, messy data: sales agents, retrieval pipelines, and multi-agent reporting. I like turning ideas into efficient, scalable backends, and I like proving they work with numbers rather than vibes.
Building and evaluating AI sales agents that talk to thousands of real customers.
Worked my way up the club's technical team.
AI agents that parse CVs and job descriptions into structured profiles, score candidate–job matches with reasoning and gaps, and write interview invites. FastAPI + PostgreSQL + SQLAlchemy backend with JWT auth, Pydantic-validated agents on Groq, and a React front end.
Semantic chunking, a spaCy-built knowledge graph, Louvain communities with LLM-written summaries, and dual local + global graph retrieval feeding a local Ollama model. Runs fully offline.
A scalar reverse-mode autodiff engine and a small neural-net library on top: configurable activations, He/Xavier init, MSE & BCE losses, and an SGD optimizer. Inspired by Andrej Karpathy's micrograd, extended into a tiny framework.
A decoder-only, character-level Transformer (6 blocks, 6 heads, 384-dim, 256 context) trained on an A100 to write Shakespeare-flavoured text. Causal self-attention, residuals and layer norm, all written out by hand, following Andrej Karpathy's GPT-from-scratch tutorial.
EDA, preprocessing and a stepwise model climb on the UCI dataset: tuned Random Forest (95.9%) → XGBoost / LightGBM (96.6%) → a soft-voting ensemble at 96.9% accuracy.
Repaints a photo in the artistic style of another image while keeping its original content. A pretrained CNN (VGG19) pulls out the shapes from one image and the textures and colors from the other, then blends them.
