A comprehensive guide to building and orchestrating agents that reason, plan, and act using foundational design patterns (Evaluator-Optimizer, Context-Augmentation, Prompt-Chaining, Parallelization, Routing, and Orchestrator-Workers).

MCP is an open protocol for connecting LLM applications with external data sources, tools, and systems. This project implements an MCP Server to easily perform retrieval and analytics on news articles with frontier models.

LLM-based applications face critical security challenges in form of prompt injections and jailbreaks. This project dives into the key architectural improvements underpinning ModernBERT, and demonstrates how to implement fine-tuning for discriminating malicious prompts. Our model closely approximates the performance of Claude 3.7 and Gemini Flash 2.0 on a mixed benchmark (NotInject, BIPIA, Wildguard-Benign, and PINT), while maintaining low latency (<40ms).

As reported by the FBI IC3, digital scams inflict devastating impacts on our society. In this project, developed for RDI Berkeley LLM-Agents Course (CS294/194-196), I built a multi-agent system with AutoGen to help users identify scam attempts. This agentic solution achieved higher accuracy compared to a prompt baseline (88.3% vs. 69.5%).

My personal top takeaways after attending TED.AI in Vienna (October 17-19, 2024) and participating in the official TED.AI Hackathon AI for Good @ UNIDO.

The ever-growing volume of research publications necessitates efficient methods for structuring such knowledge. This project implements an end-to-end automated solution using Machine Learning (UMAP, HDBSCAN), Embedding Quantization, Prompt Engineering, and an LLM pipeline to classify a dataset of 25,000 arXiv publications under a novel taxonomy of classes.

Python from-scratch implementation of the essential modules required to build and train a neural network that classifies garment images, incl. linear and flatten transformations, non-linearity, regularization, gradient descent optimizer, backpropagation, mini-batch training, and evaluation.