Born at the UC Berkeley AI Hackathon, Study Buddy evolved from a document-aware chatbot to a fully immersive AI tutor. Users upload materials, trigger hybrid semantic search, and receive streaming responses from a fine-tuned Gemma2:9b model. I integrated Hume AI's Empathic Voice Interface, enabling spoken responses that adapt tone based on emotional context. Learners don't just read; they feel the explanations.

At CalHacks 11.0, I built a platform linking specialized AI agents with LangGraph to automate the content lifecycle. It crafts content, schedules posts with LinkedIn insights, and tracks metrics. I fine-tuned LLaMA3 8B with LoRA to ensure relevance, showing how AI agents can collaborate like a real team.

Fillr automates job applications using browser automation and AI (Gemini) to generate tailored answers for open-ended questions. It includes guardrails like rate limiting and intelligent error handling. With queue management and manual review flags, Fillr combines speed and accuracy.

Built for the Llama 3 Hackathon, InsureSearch is an AI assistant that helps students understand complex insurance policies. It supports both default university documents and user-uploaded files, using Llama3 and advanced NLP to deliver precise, document-based responses. The solution is especially helpful for international students and demonstrates versatile applications in law and finance.