Farid Saud

A webcam-powered, gesture-controlled resume viewer using MediaPipe hand tracking. Navigate resume sections by tilting your hand as a dial, expand cards with an open palm, and collapse with a fist. Built with vanilla JavaScript and HTML5 Canvas - no frameworks, no build step. Try it live at fsaudm.github.io/hand.

A tool for extracting structured information from unstructured text documents (PDFs, papers) using open-source LLMs deployed on institutional GPU clusters. Supports user-defined extraction instructions and ensures all extracted data has verifiable existence in the source document. Built with LangChain and locally hosted models.

Centralized institutional knowledge system spanning 70+ DSRS repositories. Prototypes state-of-the-art RAG workflows using pgvector, FAISS, and LangChain/LlamaIndex for semantic search and retrieval across codebases, documentation, and research artifacts.

Prototype caching layer to reduce redundant LLM API calls and cost. Uses embedding-based similarity matching to serve cached responses for semantically equivalent queries, significantly reducing external API dependence.

Deployment and optimization of open-source AI models (LLM/VLM, embeddings, TTS, image generation) on A100/H100-class GPUs using vLLM, llama.cpp, and HuggingFace. Includes single-GPU and multi-GPU/distributed inference configurations via OpenAI-compatible endpoints for scalable batch processing.

Production workflows for 100K+ structured API calls to OpenAI models for research data processing. Includes robust error handling, rate limiting, structured output parsing, and cost optimization for large-scale research workflows.

Delivered a real-time demonstration of an agent-based AI image generation system combining proprietary and locally hosted models. Showcased practical LLM deployment and application for research workflows at the UIUC Ashby Prize hackathon.

Implemented object detection in construction sites using YOLOv10 and RT-DeTr architectures with PyTorch and Ultralytics. Compared performance in terms of speed and accuracy on AWS SageMaker with GPUs, using the SODA construction dataset.

Exploration of Nvidia’s Neural Infrastructure Modules (NIMs) with Llama 3.1 (8B and 405B) for reasoning and multi-lingual queries, plus StabilityAI’s Stable Diffusion XL for text-to-image generation.

Implementation of a Gaussian Mixture Model using EM algorithm, and a Hidden Markov Model through Baum-Welch and Viterbi algorithms. Detailed R markdown walkthrough available here.

Implementation of a Linear SVM using the Pegasos algorithm (Shalev-Shwartz et al., 2011). R markdown with details here.

Data analysis of historical sales from 45 Walmart stores with Robust Linear Regression on SVD-smoothed data, forecasting future sales and identifying top/bottom performing stores and departments.

Replication and extension of Professor Hadi Meidani’s Kaczmarz algorithm for real-time, short-term traffic prediction. Explored spatial vs. time-based ordering of multivariate vectors, finding spatial ordering achieved superior prediction accuracy for extreme weather conditions.

Fully-connected neural network from scratch using only NumPy, trained on MNIST (94% accuracy). Four-layer architecture with 164K parameters. All activation functions, forward/backward propagation implemented from scratch. Integrated with Weights & Biases for experiment tracking.