Muhammad Assad Ullah

Edge AI Research Engineer | Mobile Systems Specialist
Edge Intelligence Lab, Duke Kunshan University (Remote)
Department of Computer Science, University of Engineering and Technology (UET) Taxila, Pakistan
CV (PDF) / Email / ORCID / GitHub / LinkedIn / pub.dev / Portfolio
Muhammad Assad Ullah

I am a final-year Computer Science undergraduate at UET Taxila, Pakistan (expected graduation: July 2027), working remotely as an Edge AI Research Engineer at the Edge Intelligence Lab, Duke Kunshan University, under the supervision of Associate Professor Bing Luo.

My research focuses on on-device LLM inference, mobile systems engineering, and edge-AI reliability. Deploying deep learning models on consumer phones presents severe challenges: volatile system memory constraints, aggressive OS thermal throttling, and sensor corruption. My work addresses these bottlenecks by building native C/C++ bindings (JNI, dart:ffi), configuring hardware-accelerated Vulkan GPU kernels, tuning tensor memory layouts, and constructing fault-injection pipelines to diagnose runtime failure modes. My open-source testing driver has been integrated upstream into uiautomator2 and adopted by the FSE group at East China Normal University.

For my undergraduate thesis, I am developing AURAF (Adaptive Unified Runtime Architecture for Flutter), a hardware-aware on-device execution engine that dynamically coordinates token budgeting and cloud/edge fallback under physical memory and thermal constraints.

News

Oct 2026 Deposited three preprints on Zenodo covering joint fault injection, empirical failure mode analysis, and runtime observability for on-device AI (DOI: 23262270, DOI: 23250418, DOI: 23248718). SATE AI draft prepared for submission to JOSS and arXiv.
Aug 2026 Joined Duke Kunshan University's Edge Intelligence Lab as a remote Edge AI Research Engineer under Prof. Bing Luo, working on on-device fine-tuning for Vision-Language Models.
Sep 2026 Released SATE AI v1.0.0 on pub.dev, achieving a 160/160 maximum package score with 11 fault injectors and 8 model adapters across ONNX, TFLite, and llama.cpp.
Nov 2025 Our Flutter automation driver, u2_flutter, was officially accepted and listed in openatx/uiautomator2 (7.8k+ stars, PR #1196), and adopted into the Kea2 benchmark by the FSE group at East China Normal University.
2025 Awarded 1st Place at Techinance CyberHack 2025, judged by representatives from Amazon and the University of Alabama.
2024 Selected for and graduated from the Aspire Leaders Program, founded by Harvard Business School faculty.

Publications & Manuscripts

Characterising On-Device AI Failure Modes Under Joint Fault Injection and Runtime Observability
Muhammad Assad Ullah
Zenodo, 2026. DOI: 10.5281/zenodo.23262270.
[DOI] [Zenodo] [Code]
Abstract: Deploying artificial intelligence models directly on edge devices exposes machine learning runtimes to hardware volatile environments, including thermal limits, volatile system memory pressure, and corrupted input data. While previous work evaluates software-level fault injection and runtime observability in isolation, no empirical study has applied fault injection and high-frequency device tracing in a combined pipeline across heterogeneous model architectures. We present a joint empirical study combining SATE AI (software-level fault injection) and EdgePulse (non-intrusive runtime observability) on a physical Android handset (Tecno CH7n, MediaTek Helio G35, Android 12). Across a benchmark suite of 280 real-device inference traces covering three model runtimes (TensorFlow Lite, ONNX Runtime, and llama.cpp GGUF) and four execution scenarios, we establish three principal findings: (1) thermal stress causes extreme latency degradation (+486.8% for MobileNetV3 TFLite, +406.6% for ResNet-18 ONNX), while standard OS thermal APIs report nominal across 100% of stress runs; (2) malformed input injection produces a bimodal failure signature in quantized LLMs, entering 20-second looping generation cycles; and (3) vision models and edge LLMs exhibit inverted fault-tolerance profiles.
[Copy BibTeX]
@article{assadullah2026characterising,
  title     = {Characterising On-Device AI Failure Modes Under Joint Fault Injection and Runtime Observability},
  author    = {Assad Ullah, Muhammad},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.23262270},
  url       = {https://doi.org/10.5281/zenodo.23262270}
}
SATE AI: A Fault Injection and Reliability Engineering Framework for On-Device AI Models
Muhammad Assad Ullah
Zenodo, 2026. DOI: 10.5281/zenodo.23250418. Draft prepared for submission to JOSS and arXiv.
[DOI] [Zenodo] [pub.dev] [GitHub]
Abstract: On-device artificial intelligence deployment in mobile applications confronts machine learning models with volatile hardware environments: unpredictable memory pressure, dynamic thermal constraints, and corrupted sensor inputs. Cloud-hosted systems enjoy active cooling, stable resource budgets, and established chaos engineering toolchains; mobile edge AI has none of these guarantees. We present SATE AI, an open-source fault injection and reliability engineering framework designed for on-device machine learning models in mobile and desktop cross-platform applications built with Dart and Flutter. SATE AI provides a unified architecture comprising 11 domain-specific fault injectors covering process memory pressure, thermal throttling simulation, malformed tensor inputs, quantization drift, data corruption, network latency drops, model swaps, version mismatches, GPU memory pressure, and confidence threshold degradation paired with 8 model runtime adapters (ONNX Runtime, TensorFlow Lite, llama.cpp, CoreML, MediaPipe, ML Kit, MockAdapter, and custom interfaces). Validated through a 246-case automated test suite.
[Copy BibTeX]
@software{assadullah2026sateai,
  title     = {SATE AI: A Fault Injection and Reliability Engineering Framework for On-Device AI Models in Mobile Applications},
  author    = {Assad Ullah, Muhammad},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.23250418},
  url       = {https://doi.org/10.5281/zenodo.23250418}
}
EdgePulse: A Runtime Observability Framework for Quantized Large Language Models on Consumer Edge Devices
Muhammad Assad Ullah
Zenodo, 2026. DOI: 10.5281/zenodo.23248718.
[DOI] [Zenodo] [pub.dev] [GitHub]
Abstract: On-device AI execution lacks unified runtime observability across mobile platforms. Existing tools focus on isolated model profiling or single-engine metrics and give no synchronized view of memory, thermal state, battery draw, and inference latency during real workloads. We present EdgePulse, an open-source runtime observability framework capturing structured InferenceTrace telemetry on Android and iOS via native platform channels: resident memory via Debug.getPss(), OS thermal state via PowerManager.currentThermalStatus, battery current draw via BatteryManager.CURRENT_NOW, CPU utilization from /proc/stat, and per-inference latency. Analyzing 90 baseline traces and 60 sustained-load traces demonstrates that standard OS thermal APIs can miss active hardware throttling on budget ARM SoCs, making application-level timing essential for edge AI deployments.
[Copy BibTeX]
@article{assadullah2026edgepulse,
  title     = {EdgePulse: A Runtime Observability Framework for Quantized Large Language Models on Consumer Edge Devices},
  author    = {Assad Ullah, Muhammad},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.23248718},
  url       = {https://doi.org/10.5281/zenodo.23248718}
}

Systems & Research Software

SATE AI: Fault-Injection Framework for On-Device AI Models [pub.dev (160/160) / Paper]
Stress-tests mobile AI models under memory pressure, corrupted inputs, thermal throttling, and latency drops. Ships 11 domain-specific fault injectors and 8 model runtime adapters (ONNX, TFLite, llama.cpp, MediaPipe) with a 246-case automated test suite.
EdgePulse: Runtime Observability SDK for Edge LLMs [pub.dev / Paper]
High-frequency native telemetry capturing resident memory (PSS), OS thermal states, battery current draw, and CPU load per inference trace on Android and iOS. Discovered Android OS thermal API blind spot under sustained LLM token decoding.
Offline Multimodal Vision Assistant [GitHub]
Fully local Android application running Qwen3-VL-2B completely offline. Engineered custom llama.cpp bridges with Vulkan GPU acceleration, JNI and dart:ffi bindings, and tuned memory layouts for dense visual patch embeddings within mobile memory budgets.
u2_flutter: Flutter UI Testing Driver [PyPI / openatx/uiautomator2 PR #1196]
Extended uiautomator2 (7,800+ GitHub stars) to support automated UI testing on Flutter applications. Officially listed as a supported plugin in openatx/uiautomator2 and adopted by the FSE group at East China Normal University (Kea2 framework PRs #229 to #234).
AURAF Runtime Architecture (Undergraduate Thesis) [GitHub Monorepo]
Adaptive Unified Runtime Architecture for Flutter: dynamic token budgeting, physical memory and thermal monitoring, and hardware-aware cloud and edge fallback execution for cross-platform mobile apps.
AndroTest24 Extension (ASE 2024 Artifact) [GitHub Fork]
Added Flutter support to the AndroTest24 automated GUI testing benchmark suite. PR #2 merged into the official artifact repository; PR #4 under review.

Published packages on pub.dev: sate_ai, edgepulse, edgepulse_core, edgepulse_cli, offline_sync_flutter, flutter_form_guard.

Research & Work Experience

Duke Kunshan University, Edge Intelligence Lab Aug 2026 to Present
Edge AI Research Engineer (Remote) | Supervisor: Assoc. Prof. Bing Luo
  • Constructed a fully on-device Android fine-tuning pipeline for Qwen3.5-2B VLM using PEFT/LoRA; forward, backward, and weight update passes execute on consumer Android phones without cloud dependencies.
  • Engineered JNI and dart:ffi native bindings into llama.cpp and libmtmd for multimodal inference with manual tensor memory layout management and Vulkan GPU acceleration.
  • Integrated the SATE AI fault-injection framework into lab evaluation pipelines to evaluate model resilience under thermal throttling and RAM saturation.
DevCrew I/O Jul 2026 to Aug 2026
Trainee Artificial Intelligence Engineer | Lahore, Pakistan
  • Engineered production RAG pipelines connecting LLMs with vector stores and REST endpoints.
  • Handled machine learning model evaluation, dataset preprocessing, and latency benchmarking.
DevCrew I/O Jun 2025 to Aug 2025
Trainee Software Development Engineer | Lahore, Pakistan
  • Shipped features across client Flutter applications following an end-to-end mobile development lifecycle.
  • Wrote automated unit test suites, integrated RESTful APIs, and participated in Agile sprints.

Education

University of Engineering and Technology (UET) Taxila Aug 2023 to Jul 2027 (Expected)
Bachelor of Science in Computer Science | CGPA: 3.31 / 4.00
Punjab Group of Colleges, D.G. Khan Aug 2021 to May 2023
Intermediate in Computer Science (ICS) | Score: 1001 / 1100
  • Awarded 3rd Position overall in the D.G. Khan Board examination (2023).

Honors & Awards

Academic References

Prof. Bing Luo
Current Research Supervisor
Associate Professor, Edge Intelligence Lab
Duke Kunshan University, Kunshan, China
Email: bing.luo@dukekunshan.edu.cn
Dr. Aun Irtaza
Academic Supervisor
Chairman and Professor, Department of Computer Science
UET Taxila, Pakistan
Email: aun.irtaza@uettaxila.edu.pk
Dr. Muhammad Javed Iqbal
Academic Faculty
Assistant Professor, Department of Computer Science
UET Taxila, Pakistan
Email: javed.iqbal@uettaxila.edu.pk