A

Ayman Mahfuz

Computer Science at UT Austin ’27

Software engineering · systems · ML infrastructure

Experience Projects

About

Ayman Mahfuz

I’m a computer science student at The University of Texas at Austin, graduating in Spring 2027. I’m interested in foundation models, reasoning, interpretability, distributed systems, computer architecture, and the infrastructure behind modern AI.

Experience

May 2025 — Aug 2026

Arm

Platform Validation Intern → Co-op · ML Infrastructure

Built a Bayesian-optimization system guided by PMU feedback that reduced CPU and memory validation runs by more than 100× while finding top-0.2% stress configurations.

  • Reduced validation runs by more than 100× while identifying stress configurations in the top 0.2% of the search space.
  • Deployed the optimizer across three CPU-platform validation phases using cache-miss rate, memory bandwidth, and stall rate.
  • Earned Arm patent-filing approval for the method and presented production results to SVP and director leadership.

Jan 2025 — Jul 2026

Texas Robotics & AI Lab

Undergraduate Research Assistant · Robotics RL & LLM Evaluation

Built hierarchical multi-agent policies for RoboCup and a multi-turn evaluation harness for testing an RL-trained advisor’s transfer to persuasion dialogues.

  • Built C++ and Python training environments within a 400K+ line soccer stack.
  • Accelerated policy-training iterations by 2–3× and scaled experiments to more than 5 million episodes.
  • Served Qwen models with vLLM on TACC GPUs and recorded 54 full dialogue trajectories with seeded runs, fixed test splits, model identities, and run manifests.

Jan 2025 — Mar 2025

The Sunwater Institute

Data Engineer Intern

Built data and LLM pipelines for the Legis-1 platform, processing millions of legislative documents and more than 500K legal records.

  • Optimized retrieval speed, storage efficiency, and AI-readiness for a legislative database with millions of documents.
  • Developed RAG, embedding, and scalable-processing pipelines across more than 500K legal records.

Aug 2023 — May 2025

Center for Media Engagement

Software Engineer · Research Assistant

Built concurrent Selenium and REST workers that ingested 50M+ articles and 70M+ comments into 250M+ BigQuery rows.

  • Recovered transient scraper and API failures with concurrent workers and retry handling.
  • Preserved round-robin game assignments for 1,000+ participants with deterministic logic, event deduplication, and structured logging.
  • Reached 99% accuracy across clickbait, entity, sentiment, and story tasks by fine-tuning BERT models.

Aug 2023 — Jan 2025

Oden Institute

Undergraduate Research Assistant · Scientific ML Infrastructure

Improved MRI-segmentation Dice score by 12 percentage points across five-fold validation on 1,000+ scans using NVIDIA H100 GPUs.

  • Benchmarked convolutional and transformer architectures across five-fold validation on more than 1,000 scans.
  • Stabilized long-running PyTorch training by resolving GPU-memory, I/O, and mixed-precision bottlenecks in an Apptainer and SLURM pipeline.

Feb 2024 — Jan 2025

UT School of Information

Research Assistant

Studied diagnostic consistency in multi-agent LLM medical reasoning and presented the work at the UT AI Health Conference.

  • Tested consistency under demographic and symptom variations and analyzed inter-agent communication patterns.
  • Applied Cohen’s kappa, chi-square tests, and logistic regression to assess agreement, accuracy, and bias.

Jun 2023 — May 2024

University of Maryland

Research Intern

Built an NLP-driven chatbot for online news engagement and published the research at CHI 2024.

  • Led chatbot development and studies on human-chatbot dynamics for news readers.
  • Analyzed text with part-of-speech tagging, LIWC, and clustering over sentence embeddings.

Jun 2022 — Oct 2022

Lockheed Martin

Software Engineer Intern

Optimized CRM workflows with JavaScript and robotic process automation to improve device-data integrity.

  • Developed CRM workflows that centralized device data for enterprise teams.
  • Implemented robotic process automation for data deduplication.

Jun 2021 — Aug 2021

City of Austin

Software Engineer Intern

Improved post-COVID loan-processing workflows for small businesses using Python scripting and data visualization.

  • Used Python to streamline parts of the small-business loan-processing workflow.
  • Built data visualizations to support program operations.

Jun 2021 — Aug 2021

AT&T

Summer Learning Academy

Completed a summer program covering AI, business strategy, and professional development as its youngest participant.

  • Studied AI, business strategy, and professional development.
  • Collaborated on technology-focused initiatives.

Education

The University of Texas at Austin

B.S. Computer Science

Artificial Intelligence and Machine Learning

Expected May 2027

Relevant coursework

Data Structures · Algorithms · High Performance Computing · Natural Language Processing · Generative Visual Computing · Energy-Efficient Computing

Projects

01

Helm

AI-first project management with natural-language workflows, hybrid retrieval, and live meeting capture.

Live project ↗

02

Modern LLM

A 253M-parameter transformer trained from scratch through pretraining, SFT, DPO, and evaluation.

Repository ↗

03

Diffusion Models from Scratch

DDPM and DDIM implementations with custom U-Nets, attention, EMA, and sampling analysis.

Repository ↗

04

Multimodal Video Auto-Sorter

CLIP- and Whisper-based video organization with about 90% classification accuracy.

Repository ↗

05

Activation Checkpoint Integrity in PyTorch

Built a fail-fast guard for silent divergence between original and recomputed activations during backward; a controlled 40M-parameter GPU run localized identical-loss execution to 73 divergent gradient tensors before any optimizer state changed.

Publications

2024

Towards Designing a Question-Answering Chatbot for Online News: Understanding Questions and Perspectives

CHI 2024 · ACM

Md Naimul Hoque, Ayman Mahfuz, Mayukha Kindi, Naeemul Hassan.

Interviews with journalists and a reader study shaped design guidance for question-answering chatbots in newsrooms.

Read paper ↗

Submission

Natural-Language Advising for Sequential Decision Making: Learning to Persuade LLM Agents

Submitted to AAMAS 2027 · OpenReview

A study of natural-language advising in sequential decision making and persuasion with LLM agents.

View submission ↗

Skills

Languages
Python · C++ · Java · JavaScript · C · SQL
ML
PyTorch · Transformers · Reinforcement Learning · Scikit-learn
Systems
Linux · SLURM · Docker · Git · AWS · Google Cloud
Backend & Data
Node.js · Flask · Django · PostgreSQL · BigQuery