Arham Khan

Reduced our LLM‐based fraud detection system’s operating expenses by 80% via context compression and model selection. Developed a custom tokenizer, specialized for transaction data, that reduced sequence lengths for our fine‐tuned fraud detection LLM’s training and inference by 45% compared to the model’s original BPE tokenizer. This translates to larger batch sizes during training and inference. Built a foundational interpretability module for our fraud detection LLMs, which enabled us to attribute model predictions to specific tokens in the input sequence.

Developed proprietary PyTorch infrastructure for ROCm devices.

Machine Learning Engineer charged with building compiler infrastructure for PyTorch programs in tandem with efforts to develop our highly efficient ML runtime: SHARK. Development scope included contributions to prominent open-source projects such as torch-mlir and two prospective compiler frontends for PyTorch into the MLIR ecosystem: PI and SHARK-Turbine. Developed PI: a Python and C++-based importer from PyTorch to torch-mlir with a knack for tracing functional PyTorch programs. Finally achieving a pass rate in torch-mlir's end-to-end test suite of 81%, compared to an initial 68% (a 20% improvement in test coverage). Developed SHARK-Turbine: a groundbreaking pure Python importer from PyTorch to torch-mlir that translates FX Graphs captured by Torch Dynamo into torch-mlir modules and tightly integrates with OpenXLA's IREE project for targeted code generation. Evaluation upon a large set (20,000) of real-world PyTorch modules shows that SHARK-Turbine successfully imports 92% of the most popular PyTorch modules spanning Vision, Language, and Scientific Computing domains.

Was jointly awarded the Donovan J. Welch Research Fellowship from the University of Florida alongside a partner to pursue research in the development of Deep Convolutional Networks for Hand Pose Estimation on prosthetics. We were awarded $2000 to develop a mobile application that aims to provide an accessible solution to the issues that prosthetic users face when interfacing with computers. Prosthetic users often lack fine motor control and struggle to use existing computer peripherals as a result. Our project solves this by adapting hand pose models to track prostheses so that users can use crude gestures to control their device. This does not require the user to install any software or purchase any hardware - users only need a smartphone with Bluetooth capability. As there is a lack of image data for prostheses, we explore various methods to adapt existing hand pose detection models to the task including calculating an image transformation on the input that produces desirable activations in feature space, holding deep convolutional features static while retraining a series of input layers with limited data, and transfer learning. We produced and published a dataset of annotated prosthetic hands to inspire further research in this direction.

Digital Broadcast Inc. provides a suite of broadcasting and digital asset management solutions, allowing businesses to manage and distribute their media. I worked to implement the newest iteration of Digital Broadcast’s product offering: creating an automated transcription service using the IBM Watson API and implementing a new media segmentation process in their broadcast control software - Mediafire - to adhere to updated broadcasting network protocols in C++. Worked one-on-one with client representatives to create flexible media processing frameworks. Created a GUI using Python to facilitate seamless customization of the processing pipeline by distributing and modularizing the workflows for scheduled content distribution, format conversion, and media segmentation.

Anomaly detection using autoencoders: Work under Paul Gader. We investigated employing autoencoders as anomaly detectors for image data. Specifically, we trained autoencoders on popular datasets such as MNIST, FashionMNIST, and CIFAR-10 and then collected statistics on reconstruction error. Furthermore, we clustered the embeddings from the encoder. Reconstruction error and minimum cluster distance of each sample were compared against threshold values to identify anomalous data points. We showed promising results for anomaly detection, especially in addressing adversarial examples. In the future, we plan to combine embedding information with extracted image features to improve classification accuracy in adversarial environments. Image Super-resolution using Deep Learning. Work under Paul Gader and Joseph Wilson. We studied DNNs for super-resolution for the purposes of improving detector accuracy in the context of a lack of data. We compared deep residual networks, spatial pyramid networks, filter approaches such as the Adaptive Wiener Filter, and classic upsampling approaches such as interpolation along dimensions of both reconstruction fidelity and classifier accuracy improvements. Results showed modest improvements in classifier accuracy, particularly for very low-resolution samples. We aim to improve upon these methods in the future by leveraging features from a store of similar images to inform our super-resolution models.

TapToBook provides businesses with marketing automation tools and analytics. Implemented an email templating system using AngularJS that allowed clients to design and send mass marketing emails to subscribers through the TapToBook application interface, enabling clients to easily produce scalable email campaigns. Created an accompanying set of SQL scripts that allowed for processing and storage of client marketing templates. Built an in-browser photo editor using Javascript and AngularJS.

Unsupervised Classification of Natural Language Documents: In collaboration with Evan Seamone and Melissa Whiley. We analyzed semi-structured natural language documents collected from Veteran’s Affairs compensation cases in order to facilitate a statistical analysis of judicial outcomes for traumatic discrimination cases involving racial or sexual abuse. We developed several unsupervised machine learning approaches including applying Bag-of-Words and TF-IDF vector space representations of documents and implementing KMeans and Latent Dirichlet Allocation in order to cluster documents according to content, ultimately reducing the search space from 120,000 documents to a final set of 654 documents. We demonstrated that victims of traumatic abuse in the military are consistently denied compensation due to a lack of evidence or administrative barriers.