Harsh Karia

Incoming intern at Gemini’s OnChain team working on AI infrastructure, prediction markets and security

• Researching LLM memory behavior and data persistence under guidance from Prof. Zubair Shafiq, with a focus on developing first-of-its-kind benchmarks for data persistence, involuntary personalization, safe forgetting, and long-term memory evaluation • Conducting research into memory of LLMs to identify lapses in data deletion, involuntary personalization, vulnerabilities, and commercial memory controls in C++ and Python from dataset of 200,000 conversations • Previously led analysis of agentic computer use tools, using MITMproxy, HAR logs, and custom network scrapers to uncover hidden data flows and privacy risks • Collaborating on tools and metrics that quantify model forgetfulness, hidden personalization drift, and residual exposure, contributing to the development of safer, more controllable foundation models

• Focused on code security, cryptocurrency and blockchain infrastructure and augmenting manual code reviews of 1 million+ lines of code using C++, Python, Go, and Rust and testing with PyTest • Creating custom abstract syntax trees + RAG system to represent syntactic structure of code and using machine learning for fuzzy matching with consensus layer specs to identify vulnerabilities 80% faster

• Developing a web-based AI assistant for behavioral and technical interview prep, using Next.js, TypeScript, and Supabase • Integrated Hume AI’s computer vision models to detect facial expressions and emotional shifts in real time, generating personalized feedback on delivery, tone, and clarity • Building LLM-based question generation and feedback flows using Mistral and prompt engineering to simulate interviewer behavior • Deployed application on Modal, with scalable back-end APIs and secure session management for multi-user interviews

• Collaborating with UC Davis Men’s and Women’s Tennis teams to automate match analysis and reduce coaching film review time by 90% • Built a full-stack platform for coaches to upload film and receive game statistics, shot classifications, and heat maps generated by in-house ML models • Trained custom models for court detection, ball tracking, bounce detection, and shot selection using PyTorch, Detectron2, and OpenCV, leveraging NVIDIA A100 GPUs from Modal for cloud based inference • Integrated pose estimation and shot classification to analyze opponent patterns, enhancing strategic planning and post-match reviews

• My free intro to CS courses on Udemy and have amassed over 8000 students • During Covid-19 I struggled to keep up with my studies and so I started building out introductory courses to learn and decided to share them with the world • Inspired by this success I am currently working on building out longer courses for full stack development, agentic workflows, LLM development, and security

• As Technical Lead of cybersecurity team for UC Davis Google Developers Student Club, created full stack password management app in React.js and used deep learning in Python to create a facial recognition system to view passwords, placing 1st in beginner projects category during winter competition • Developed cryptocurrency using blockchain technology and RSA encryption with unique keys and allowed for users to buy, sell, and exchange their coins, placing 1st in advanced projects category in spring competition • Led AI/ML Team to develop models for sports betting applications using PyTorch, React + TypeScript (for user interface), Pandas, NumPy

• Directed a team of 4 in software planning and scheduling for UC Davis’ CITRIS Aviation Competition Team to design an optimized system for 32 air taxis • Engineered prototypes of a multifunctional application enabling seamless interaction between passengers and vehicle operators within the new air mobility system, enhancing scheduling efficiency using Python and JavaScript

• Introduced automations for insurance data parsing and email marketing using OpenAI’s LLM models and building internal APIs for financial data modelling, reducing resource requirements by 100x • Built out insurance policy pricing algorithm using large self curated datasets and NLP techniques

• Spearheaded end-to-end development of a mobile application using React Native and JavaScript, facilitating data collection for a nonprofit and successfully launching the app on the App Store • Designed and deployed frontend user interfaces, integrating object detection models with 95% accuracy to automate wild horse identification in images