I'm Vishal, a graduate student at UC Santa Cruz pursuing an MS in Natural Language Processing, based in California. I work across AI systems, ML infrastructure, and NLP research — with a focus on RAG pipelines, LLM evaluation, agentic AI, and production ML systems. My background spans industry and research (GUII Lab, IEEE Best Paper 2024).

• CriticRAG
  Multi-agent RAG pipeline built from scratch — no LangChain, no LlamaIndex. Features MoE retrieval routing (dense/sparse/hybrid), a critic-driven validation loop with LLM-as-judge scoring, self-query reformulation, tool calling (live Wikipedia + calculator), and self-consistency generation. Evaluated on HotpotQA distractor split: +2% Exact Match and +2.5% Token F1 over single-pass RAG baseline across 100 hard multi-hop questions.

• Integer Recognition Using Neural Networks
  Developed a custom neural network in Java for real-time integer recognition, featuring custom weight and bias handling. The model efficiently processes low-resolution and stylized images using image preprocessing pipelines, enabling accurate recognition even in challenging conditions.

• Flappy Bird Agent with Reinforcement Learning and Neuroevolution
  Created an AI agent for the Flappy Bird game using NeuroEvolution of Augmenting Topologies (NEAT) and Q-learning. This agent learns through fitness-based evolution and adapts dynamically to the game environment, improving its survival rate with each generation.

• Self-Driving Car Simulation
  Designed an autonomous car simulation using the NEAT algorithm and Python, which visualizes adaptive decision-making for navigating a 2D track. This project simulates real-time sensor feedback and optimizes collision avoidance through an evolving neural network.

• Fake News Detection using Passive-Aggressive Classifiers
  Developed a real-time fake news detection model leveraging TF-IDF vectorization and a passive-aggressive classifier. This web application classifies news articles as "Fake" or "Real," providing a practical solution to counter misinformation.

• Deep Learning Upscaling Model for Textures
  Created a deep learning model using Residual in Residual Dense Blocks (RRDB) to upscale low-resolution game textures by up to 4x. The model enhances image clarity and detail, with efficient GPU processing using PyTorch, suitable for real-time game development.