KM Khalid Saifullah · Cleveland, OH

Tasin

I find the problem by talking to the people living with it, then build the fix and run it in production.

I build the things I've watched people quietly route around. Most recently that meant AtlasDesk, an IT service platform I built solo at the college where I was also the one answering the tickets. It replaced a $22K/year vendor tool for 4,705 people, and I now lead the four college engineers who maintain it.

LinkedIn GitHub

Open to product, engineering, and IT roles.

Selected work

Systems I designed, shipped, and now run.

Experience

Read it from the product side, the engineering side, or both.

  1. ITSM platform in production at The College of Wooster. Built it solo; now lead the 4 college engineers who maintain it.

    • Interviewed 38 users (4 directors, 14 technicians, 20 student techs) and found the vendor service desk couldn’t link tickets to devices, track loaners, search ticket text, or run a report.
    • Built the MVP in 3 days and won unanimous buy-in from 4 divisional directors. Owned the post-launch roadmap, prioritized by ROI.
    • Deployed it on the college’s on-prem SQL Server and Microsoft Entra tenant (SAML 2.0 SSO), replacing a $22K/year vendor tool for 4,705 users. It handled 1,600+ tickets from 143 departments in its first 6 months.
    • 34 operators rated it 8/10, against 3/10 for the old tool.
    • Migrated six years of data and 6,000+ assets out of TopDesk and scattered Excel sheets with zero records silently dropped, logging every unmatched value to an audit trail.
    • Shipped a retrieval copilot over 33,000+ resolved tickets, with personal data redacted before any model call. Median resolution time fell 72% (46.7 h to 13.1 h) since launch.
    • Built a read-only MCP server and AI agent in Microsoft Teams and Microsoft 365 Copilot that replaced engineer-written SQL reports for 4 department decision makers, and decided what it refuses: per-person stats and ticket text.
    • Ranked 50+ feature requests by ROI and shipped print-billing automation. Said no to an in-browser voicemail player: the college gets 5–6 voicemails a week, and technicians already check them at the desk phone.
    • Kept one codebase deployable to on-prem and SaaS from a single branch: a CI linter rejects PostgreSQL-only patterns, and a matrix release builds against PostgreSQL and SQL Server 2022.
    Read the case study →
  2. Scope grew from support into solution design and platform ownership.

    • Primary technical contact for 4,705 faculty, staff, and students. Resolved 500+ support tickets at 90% satisfaction and turned recurring problems into AtlasDesk features.
    • Found 95% of my 500+ resolved tickets repeated a solved issue or an existing help article, and presented it to the 4 directors so technicians could shift that time to process fixes.
    • Administer users and Intune-managed devices in the college’s Microsoft Entra ID tenant, and support both Windows and Mac users.
    • Traced a wave of campus printing tickets to Macs on macOS 27, where Apple dropped Rosetta, and fixed the root cause by installing Rosetta instead of closing tickets one by one.
    • Wrote a one-run PowerShell pipeline on OSDCloud and Intune that images a laptop, enrolls it in Intune, and leaves it login-ready on campus credentials. Per-machine setup dropped from hours to about 10 minutes across 200+ lab machines.
    • Built an evaluation framework that benchmarked 13 LLMs across single- and multi-agent architectures, with GPT-4 as judge. Its results informed model selection for Schneider’s enterprise co-pilot.
    • Prototyped an LLM co-pilot (LlamaIndex, FastAPI) for plain-English questions over 250,000+ energy-data workflows, connected to 15+ enterprise tools.
  3. Fractional real-estate investment startup.

    • Owned the backend for cart, reservation, checkout, payment, and sell-back flows from API design through the v1.0 production launch.
    • Built automated bank-transfer reconciliation: OCR’d user receipts (English and Bangla) matched against DKIM-verified bank emails, settling payments within about a minute.
    • Prevented overselling with event-based unit locks: the first reservation held units for up to 24 hours, competing requests failed immediately, and unpaid holds released automatically.
  4. HR-tech startup. Five-person team.

    • Interviewed 8 Bangladeshi companies and found hiring ran on Google Forms, email, and unstandardized walk-in tests, making screening slow and often biased by gender or name.
    • Co-designed an ATS that ranked candidates by automated assessment and résumé-to-JD scores so HR picked the top 5. Won 3 of the interviewed companies as beta clients.
    • Built the backend in Django and integrated SSLCommerz payment processing as one of two technical co-founders.

Research

Papers

Multi-Agent Coordination in Autonomous Vehicle Routing: A Simulation-Based Study

2025

Published · Research Paper · KM Khalid Saifullah

This study explores how multi-agent interaction enhances autonomous vehicle (AV) decision-making in dynamic traffic environments. While traditional AV models focus on individual autonomy, real-world traffic scenarios often require collective behavior through inter-agent communication and coordination. To investigate this, we developed a graph-based simulation environment that enables vehicle agents to exchange information and reroute in real time in response to road obstacles. Our findings demonstrate that communication and adaptive rerouting significantly reduce average wait times and improve travel efficiency. Furthermore, we introduce a lightweight memory mechanism—Object Memory Management (OMM)—which allows agents to retain knowledge of previously encountered obstacles. This feature proved critical in avoiding routing loops and redundant decisions. Together, these results highlight the potential of communication- and memory-enhanced agents in creating resilient, cooperative AV systems capable of navigating complex and unpredictable traffic networks. Read abstract

This study explores how multi-agent interaction enhances autonomous vehicle (AV) decision-making in dynamic traffic environments. While traditional AV models focus on individual autonomy, real-world traffic scenarios often require collective behavior through inter-agent communication and coordination. To investigate this, we developed a graph-based simulation environment that enables vehicle agents to exchange information and reroute in real time in response to road obstacles. Our findings demonstrate that communication and adaptive rerouting significantly reduce average wait times and improve travel efficiency. Furthermore, we introduce a lightweight memory mechanism—Object Memory Management (OMM)—which allows agents to retain knowledge of previously encountered obstacles. This feature proved critical in avoiding routing loops and redundant decisions. Together, these results highlight the potential of communication- and memory-enhanced agents in creating resilient, cooperative AV systems capable of navigating complex and unpredictable traffic networks.

Autonomous VehiclesMulti-Agent SystemsV2V CommunicationDynamic ReroutingTraffic Simulation Read paper ↗

Sentiment Analysis in Software Engineering: Evaluating Generative Pre-trained Transformers

2024

Published · National Conference on Undergraduate Research (NCUR) · KM Khalid Saifullah

Sentiment analysis plays a crucial role in understanding developer interactions, issue resolutions, and project dynamics within software engineering (SE). While traditional SE-specific sentiment analysis tools have made significant strides, they often fail to account for the nuanced and context-dependent language inherent to the domain. This study systematically evaluates the performance of bidirectional transformers, such as BERT, against generative pre-trained transformers, specifically GPT-4o-mini, in SE sentiment analysis. Using datasets from GitHub, Stack Overflow, and Jira, we benchmark the models' capabilities with fine-tuned and default configurations. The results reveal that fine-tuned GPT-4o-mini performs comparable to BERT and other bidirectional models on structured and balanced datasets like GitHub and Jira, achieving macro-averaged F1-scores of 0.93 and 0.98, respectively. However, on linguistically complex datasets with imbalanced sentiment distributions, such as Stack Overflow, the default GPT-4o-mini model exhibits superior generalization, achieving an accuracy of 85.3% compared to the fine-tuned model's 13.1%. These findings highlight the trade-offs between fine-tuning and leveraging pre-trained models for SE tasks. Read abstract

Sentiment analysis plays a crucial role in understanding developer interactions, issue resolutions, and project dynamics within software engineering (SE). While traditional SE-specific sentiment analysis tools have made significant strides, they often fail to account for the nuanced and context-dependent language inherent to the domain. This study systematically evaluates the performance of bidirectional transformers, such as BERT, against generative pre-trained transformers, specifically GPT-4o-mini, in SE sentiment analysis. Using datasets from GitHub, Stack Overflow, and Jira, we benchmark the models' capabilities with fine-tuned and default configurations. The results reveal that fine-tuned GPT-4o-mini performs comparable to BERT and other bidirectional models on structured and balanced datasets like GitHub and Jira, achieving macro-averaged F1-scores of 0.93 and 0.98, respectively. However, on linguistically complex datasets with imbalanced sentiment distributions, such as Stack Overflow, the default GPT-4o-mini model exhibits superior generalization, achieving an accuracy of 85.3% compared to the fine-tuned model's 13.1%. These findings highlight the trade-offs between fine-tuning and leveraging pre-trained models for SE tasks.

Sentiment AnalysisSoftware EngineeringTransformer ModelsGPTBERTFine-tuningNLP Read paper ↗

Education

The College of Wooster

B.A. in Computer Science & Business Economics

Departmental Honors in Computer Science

2025

Leadership

  • Google Developer Student Club

    Co-Lead, 2024–2025. 68 members; ran 4 résumé and project workshops.

  • Stanford TreeHacks 2024

    Built GovAI, a multi-agent policy council.

  • Residential Assistant

    The College of Wooster, 2022–2024. Mentored 50+ residents.