Projects
Things I've built and researched.
Voice AI Agent
Real-time multilingual voice AI agent built with LiveKit and LlamaIndex, integrated with a Proto Hologram device for an immersive business use case. Implemented a cascaded pipeline that reduced inference costs by 32% and lowered perceived latency by 10% compared to speech-to-speech solutions.
RAG Document System
Internal retrieval-augmented generation system leveraging Qdrant as a vector database, Inngest for workflow orchestration, and FastAPI for the backend. Boosted employee efficiency by 50% by enabling intelligent search and retrieval across company documents using LangChain.
Network Attack Detection
Research project combining autoencoders for dimensionality reduction with reinforcement learning to model network attack patterns on the CICIDS2017 dataset, achieving approximately 97% accuracy. Refined hyperparameters using grid search and Bayesian optimization, increasing detection recall to 92% while reducing false positives by 6%.
Wajez OCR
AI-powered OCR solution for logistics co-developed during the Masar Program. Leveraged LLM fine-tuning and PyTesseract to accurately extract and process text from shipping documents and labels, streamlining logistics workflows.
Carbon Emissions Analyzer
Emission analysis tool designed to calculate and visualize client-specific carbon emission data. Includes an AI assistant built with the OpenAI API and Streamlit that parses and analyzes internal company documents using optimized prompt engineering techniques.
Agentat
AI workforce platform co-founded during Misk Launchpad 8.0. Agentat aims to streamline enterprise operations by deploying autonomous AI agents that handle repetitive tasks, enabling teams to focus on higher-value work.