John Pangasest. 1997

Projects

Research, open source, and things that shipped.

A chronological portfolio of the work behind the papers, tools, and early builds: agentic AI, software engineering, Mozilla, machine learning, and student projects.

7

projects, kept in the same order as the working notes.

  1. 01
    September 2025/Journal publication

    The Hands Behind the Agents

    Information and Software Technology (Elsevier)

    Paper

    A large-scale empirical study of the real-world challenges developers face when building with agentic frameworks such as LangChain, CrewAI, and AutoGen.

    Areas: Agentic AI, Developer experience, LDA, Stack Overflow

    Authors: Rinad Hamid, John Pangas, Ahmad Abdellatif

    What I Did

    • Analyzed 2,658 Stack Overflow discussions related to agentic frameworks.
    • Applied LDA topic modeling and tag co-occurrence analysis to surface recurring development challenges.
    • Investigated developer question patterns, framework pain points, unanswered-post rates, and resolution times.

    Details

    • Framework configuration and tool integration were among the most common sources of frustration.
    • 71.1% of posts were procedural "How-to" questions.
    • LLM execution and runtime issues had the highest unresolved rate, with 86% remaining unanswered.
    • Vector database problems had the longest median resolution time at 100.3 hours.

    Impact

    The findings give framework maintainers and practitioners evidence for improving documentation, debugging workflows, abstractions, and developer experience in agentic AI systems.

  2. 02
    March 2025/Conference publication

    Prompting Matters

    International Conference on Software Maintenance and Evolution (ICSME)

    PaperCode

    An empirical study of how prompting strategies affect LLM-generated class-level code, moving beyond method-level generation into complete class implementations.

    Areas: LLMs, Code generation, Prompting, Software engineering

    Authors: Adam Yuen, John Pangas, Md Mainul Hasan Polash, Ahmad Abdellatif

    What I Did

    • Evaluated zero-shot, few-shot, chain-of-thought, and chain-of-thought few-shot prompting.
    • Compared performance across GPT and Llama models.
    • Assessed generated code using functional correctness, BLEU-3, ROUGE-L, readability, and maintainability metrics.
    • Conducted qualitative analysis to identify common failure patterns.

    Details

    • Context-rich prompting improved functional correctness by up to 25%.
    • BLEU-3 improved by 31%, and ROUGE-L improved by 50%.
    • Procedural logic and control-flow errors accounted for 31% of observed failures.

    Impact

    This work provides practical guidance for developers and researchers building AI-assisted software engineering tools, showing how prompt design directly affects code quality and reliability.

  3. 03
    March 2024/Industry research collaboration

    Using LLMs to Bridge the Gaps in QA Test Plans at Firefox

    IEEE Software

    Paper

    A Mozilla and University of Calgary collaboration exploring how GPT-4 can assist Firefox QA engineers by generating test cases for new browser features.

    Areas: Firefox, QA, GPT-4, Test generation

    Authors: John Pangas, Suhaib Mujahid, Ahmad Abdellatif, Marco Castelluccio

    What I Did

    • Designed and developed a GPT-4-powered test case generation pipeline.
    • Built a comparison system for generated and human-authored test cases.
    • Developed mechanisms to identify missing coverage in manually written test plans.
    • Evaluated generated test cases against those created by professional QA engineers.

    Details

    • The workflow ingests feature descriptions, testing scope, and historical test cases.
    • GPT-4 generates candidate test cases for human review.
    • The system compares generated tests against human-authored plans to identify missing scenarios and coverage gaps.

    Impact

    The project demonstrates how LLMs can augment QA workflows, helping teams create more comprehensive test plans while reducing manual effort.

  4. 04
    February 2023 - Present/Open source contributions

    Mozilla BugBug Contributions

    Repository

    Ongoing contributions to Mozilla's BugBug platform, a machine-learning-powered system used for bug triage, defect prediction, and software engineering automation.

    Areas: Mozilla, Bugzilla, Machine learning, Open source

    What I Did

    • Submitted over 25 pull requests.
    • Contributed bug fixes, enhancements, and infrastructure improvements.
    • Collaborated with Mozilla maintainers and contributors on production-grade machine learning tooling.

    Impact

    Helped improve open-source tools used by software engineering researchers and Mozilla developers to automate development workflows at scale.

  5. 05
    November 2022 - April 2023/Open source contributions

    Scikit-Learn Contributions

    Repository

    Contributions to scikit-learn, one of the world's most widely used machine learning libraries.

    Areas: Python, scikit-learn, Machine learning, Open source

    What I Did

    • Submitted 8 pull requests.
    • Fixed bugs and improved library functionality.
    • Participated in community discussions and code reviews.

    Impact

    Contributed to software used by millions of machine learning practitioners, researchers, educators, and students worldwide.

  6. 06
    December 2021 - June 2022/Bachelor's capstone project

    Deep Learning for Fire Detection in Low-Income Countries

    Zhejiang University of Technology

    A low-cost intelligent fire detection system for real-time household fire detection using computer vision, deep learning, cloud services, and embedded hardware.

    Areas: Computer vision, Raspberry Pi, AWS, Edge AI

    What I Did

    • Built a fire detection model using depthwise separable convolutional neural networks.
    • Trained on a dataset containing more than 8,000 images.
    • Achieved 94-95% fire detection accuracy.
    • Optimized inference to 15-17 FPS on Raspberry Pi hardware.
    • Integrated AWS SageMaker and AWS Pinpoint services.
    • Designed a deployment architecture costing less than $70.

    Details

    • Real-time fire detection from surveillance video.
    • Occupancy detection for people and pets.
    • Instant SMS notifications, scene snapshots, and incident reporting.
    • Raspberry Pi edge deployment.

    Impact

    Demonstrated how AI-powered safety systems can be deployed affordably in resource-constrained environments while maintaining strong performance.

    Rated Outstanding by the Computer Science Project Oversight Board at Zhejiang University of Technology.

  7. 07
    November 2021 - January 2022/Android development project

    Student To-Do List App

    An Android productivity app for helping students manage assignments, homework, and personal tasks through a simple, student-focused interface.

    Areas: Kotlin, Android, Material Design, Productivity

    What I Did

    • Built the application using Kotlin.
    • Implemented a priority-based task sorting algorithm.
    • Designed the interface using Material Design principles.
    • Created features for organizing and tracking academic workloads.

    Details

    • Task creation and management.
    • Priority-based sorting.
    • Date-based organization.
    • Student-focused user experience.

    Impact

    Provided students with a lightweight productivity tool for managing coursework and daily responsibilities.