STUDY DIRECTORY // SELECT A SHELFI. ResearchCOORDINATION · LLMs ↓II. PublicationsPAPERS & ARXIV ↓III. ProjectsTESTBEDS · ML CODE ↓IV. Experience & CVEXPERIENCE · EDUCATION ↓U.S. DEPT. OF STATE4.0 / 4.0 M.S. AI · RIT1ST RANK B.S. CS ’21FULBRIGHT SCHOLARAI/ML SDE · SUMMER ’25★ Full-Time Return OfferEAST PALO ALTO · SPARKAWS AI/ML INTERNPH.D. GRA · ROCHESTERCoordination & PlanningDECENTRALIZED TEAMSKAS LAB · RITCV.PDFOPEN ↗KHAWAJA ABAID ULLAHPH.D. GRA · RIT KAS LABPull brass chain to switch study lighting

First-year PhD student in ECE · Rochester Institute of Technology

Khawaja Abaid Ullah

I’m a first-year PhD student in Electrical and Computer Engineering at Rochester Institute of Technology, where I study multi-agent coordination and decision-making at the Khojasteh Autonomous Systems Laboratory. My current work focuses on decentralized task allocation and LLM-assisted planning for autonomous systems.

I also build machine-learning software, including Teras, a multi-backend tabular deep-learning library, and have contributed code and tests to Keras. Previously, I completed an SDE internship at AWS and earned my MS in Artificial Intelligence at RIT as a Fulbright Scholar.

01 / Research

Research Focus

Decentralized coordination today; algorithm discovery as an emerging direction.

Research overview
  1. Multi-agent coordination & decision-making

    Current research · KAS Lab

    My research focuses on how autonomous agents coordinate decisions, share information, and allocate tasks under communication and resource constraints. I am particularly interested in decentralized coordination and in combining LLM-assisted planning with explicit coordination algorithms.

    • Decentralized task allocation
    • LLM-assisted planning
    • Autonomous systems
  2. Algorithm discovery & efficient reasoning

    Emerging direction & long-term goals

    I am beginning to explore LLM-guided algorithm discovery and optimization, where candidate solutions are improved using evaluation feedback. Longer term, I want to develop compact, computationally efficient reasoning models that can run directly on robots and edge devices, helping autonomous teams adapt and coordinate reliably. Multi-robot search and rescue is one application I hope to work toward.

    • Emerging: LLM-guided algorithm discovery
    • Goal: on-device reasoning

02 / Papers

Selected Papers and Preprints

Current coordination research and contributions to classification defenses.

All papers & collaborative work

arXiv:2610.41044 · Preprint; under review · 2026

Consensus Within, Competition Between: Decentralized LLM Teams for Perimeter Defense

Khawaja Abaid Ullah, Ananya Acharya, Mohammad Javad Khojasteh

In the evaluated perimeter-defense simulations, decentralized LLM teams with external auction-based conflict resolution (CBAA) had lower breach rates than centralized or independent teams. The reproducible testbed examines communication constraints, decision authority and multimodal visual deception.

BibTeX for Consensus Within, Competition Between: Decentralized LLM Teams for Perimeter Defense
@misc{ullah2026consensus,
  title = {Consensus Within, Competition Between: Decentralized LLM Teams for Perimeter Defense},
  author = {Khawaja Abaid Ullah and Ananya Acharya and Mohammad Javad Khojasteh},
  year = {2026},
  eprint = {2610.41044},
  archivePrefix = {arXiv},
  primaryClass = {cs.AI},
  url = {https://arxiv.org/abs/2610.41044}
}

arXiv:2607.15467 · Preprint; under review, IEEE Open Journal of the Computer Society · 2026

ADS-C: Antidistillation Sampling for Classification

Khawaja Abaid Ullah, Mohammad Javad Khojasteh

An inference-time defense that changes served probability outputs while provably preserving top-1 predictions. It reduced soft-label distilled-student accuracy in the evaluated image-classification experiments; unchanged labels remain available to hard-label attackers.

BibTeX for ADS-C: Antidistillation Sampling for Classification
@misc{ullah2026adsc,
  title = {ADS-C: Antidistillation Sampling for Classification},
  author = {Khawaja Abaid Ullah and Mohammad Javad Khojasteh},
  year = {2026},
  eprint = {2607.15467},
  archivePrefix = {arXiv},
  primaryClass = {cs.LG},
  url = {https://arxiv.org/abs/2607.15467}
}

Manuscript under review · 2026

AntidistillBench: A Unified Benchmark of Test-Time Antidistillation Defenses for Classification

Khawaja Abaid Ullah, Mohammad Javad Khojasteh

Implements test-time classification defenses under a common model-extraction threat model: an attacker queries a classifier and learns a student from its outputs. Existing defenses run through a matched-cost evaluation protocol to compare the accuracy of students trained on defended outputs.

BibTeX for AntidistillBench: A Unified Benchmark of Test-Time Antidistillation Defenses for Classification
@unpublished{ullah2026antidistillbench,
  title = {AntidistillBench: A Unified Benchmark of Test-Time Antidistillation Defenses for Classification},
  author = {Khawaja Abaid Ullah and Mohammad Javad Khojasteh},
  year = {2026},
  note = {Manuscript under review}
}

03 / Projects

Projects & Code

A coordination testbed, a multi-backend ML library, merged upstream contributions, and defense code.

All projects

Built a reproducible simulation testbed comparing centralized, auction-based consensus (CBAA) and independent LLM teams. Perimeter defense is the experimental setting for studying task allocation under communication constraints, decision authority and multimodal visual deception.

  • Multi-agent coordination
  • LLM-assisted planning
  • CBAA
  • Reproducible simulation

Teras

Independent open-source library

Created a Keras 3 library implementing published tabular deep-learning architectures for classification, regression, data imputation and synthetic-data generation, with JAX, PyTorch and TensorFlow backends.

  • Keras 3
  • JAX
  • PyTorch
  • TensorFlow
  • Tabular Deep Learning

Keras contributions

Independent upstream contributions

Implemented Beta and Binomial random samplers and unit tests across JAX, PyTorch, TensorFlow and NumPy. Merged upstream in Keras 3.

  • Keras 3
  • JAX
  • PyTorch
  • TensorFlow
  • NumPy
  • Unit tests

Released a PyTorch implementation of margin-bounded antidistillation sampling and an evaluation pipeline for CIFAR-10, CIFAR-100 and Tiny-ImageNet. Related benchmark work, AntidistillBench, compares classification defenses under a unified threat model and matched-cost protocol; its manuscript is under review.

  • PyTorch
  • ML Security
  • Antidistillation
Earlier student & personal projects (5)
  • Prewordict — A word-cloud guessing game built from clusters of words in Medium articles. GitHub
  • Gamer Assistant — A CS50 project combining game recommendations, PC suggestions, and FPS prediction.
  • PROLEAs — Undergraduate work on crowd detection and behavior analysis using computer vision.
  • FPS Predictor — An early machine-learning project estimating game performance from hardware data.
  • Tweets Cloud Bot — A Twitter bot that turned recent tweets into customizable word clouds. GitHub

04 / Record

Experience & CV

Research, industry, teaching, academic service, and education.

Full experience & education

Appointments & Industry

  1. Graduate Research Assistant (Ph.D. in ECE)

    Fully funded graduate research assistantship

    Khojasteh Autonomous Systems Laboratory (KAS Lab), RIT · Rochester, NY

    • Advised by Dr. Mohammad Javad Khojasteh (expected graduation: May 2029).
    • Study multi-agent coordination and decision-making, particularly decentralized task allocation and LLM-assisted planning under communication and resource constraints.
  2. Software Development Engineer Intern, AI/ML

    Full-Time Return Offer

    Amazon Web Services (AWS) · East Palo Alto, CA

    • Engineered core Apache Spark testing and evaluation modules for an internal benchmarking platform, closing feature-parity gaps with its legacy predecessor.
    • Independently identified and delivered additional benchmarking capabilities beyond original project scope, earning team leadership recognition and a full-time return offer.
  3. Teaching Assistant — Principles of Robotics

    ROS 1 / ROS 2 / Docker

    Rochester Institute of Technology · Rochester, NY

    • Develop lecture and lab material for undergraduate and graduate robotics students; designed reproducible ROS Docker containers and VM images replacing a brittle third-party setup.
  4. Junior Python Developer (Remote)

    Comundo · Copenhagen, Denmark

    • Built semi-supervised data-curation pipelines for building energy-consumption records used in CO2 estimation and downstream forecasting.

Education

  1. Ph.D. in Electrical and Computer Engineering

    Fully funded graduate research assistantship

    Rochester Institute of Technology · Rochester, NY

    Advised by Dr. Mohammad Javad Khojasteh in the Khojasteh Autonomous Systems Laboratory.

  2. M.S. in Artificial Intelligence

    U.S. Fulbright Scholar · 4.0 / 4.0 GPA

    Rochester Institute of Technology · Rochester, NY

    Funded by the U.S. Department of State Fulbright Foreign Student Program. Thesis: “Antidistillation Sampling for Classification.”

  3. B.S. in Computer Science

    Ranked 1st in Class of 2021

    University of Narowal · Narowal, Pakistan

    Graduated first in the Computer Science class of 2021.

Academic Service

Jun 2026 – Present

IEEE Signal Processing Letters

Began reviewing manuscripts for IEEE Signal Processing Letters in June 2026.

Technical Stack

Languages
Python · C · C++ · Bash
ML & Deep Learning
JAX · Flax (Linen & NNX) · Keras 3 · PyTorch · TensorFlow · Ray RLlib · PettingZoo · NumPy · Apache Spark
Robotics & Systems
ROS 1 · ROS 2 · Docker · Linux · Git · Arduino
Research Focus
Multi-Agent Reinforcement Learning · Decentralized Task Allocation · Antidistillation & ML Security · Knowledge Distillation · LLM-Assisted Planning