Intro

Yonas Gebremedhin

I am a Software Engineer at Microsoft, solving problems at the intersection of performance, scale, and collaboration. Beyond shipping code, I contribute to AI-driven accessibility research, thriving where engineering meets experimental discovery.

Resume, Projects

LinkedIn · GitHub


Building

I engineer scalable communication apps and mobile experiences where reliability, performance, and user trust are non-negotiable.

  • I co-lead the delivery of core features for Microsoft Teams, owning the full lifecycle from technical design and cross-functional alignment to global production rollout.
  • I optimize high-traffic communication surfaces, ensuring predictable performance in environments where milliseconds of latency impact millions of concurrent users.
  • I bridge rapid feature delivery with long-term architectural health, driving modularity and reducing technical debt to keep complex codebases extensible and resilient.
  • Beyond enterprise systems, I design and ship native iOS applications in Swift, applying the same principles of performance, reliability, and intuitive UX.

Projects

Quick Views (Microsoft Teams)

Streamlined how users surface critical updates in high-volume environments, reducing cognitive load and helping teams spend less time searching and more time doing.

Chats/Channels Modernization (Microsoft Teams)

Modernizing core communication experiences by improving comprehension, workflows, and system consistency at scale.


Publications and papers

A Hybrid LLM-Computer Vision Framework for Automated Structural Remediation of Complex Educational Digital Assets

Aegis-A11y is a hybrid AI framework for automating accessibility remediation of complex educational digital assets (e.g., STEM diagrams, multi-column layouts, nested tables, and scanned worksheets) that often fail under OCR-only or heuristic auto-tagging tools. The system follows a Decomposition–Reasoning–Reconstruction (DRR) pipeline: a computer-vision layout stage detects and classifies page regions, a multimodal LLM layer generates context-aware semantics (especially alt-text), and a deterministic verifier checks structural rules (e.g., heading order, table associations, and checklist-style WCAG screening) and triggers corrective loops when violations are found.

View paper


Now

I am always looking to connect with engineers and researchers working at the intersection of AI, Accessibility, and Scalable Systems.


Connect

Find me as @yonashailug on LinkedIn / Medium.