Kay Wijerathne

Accessibility · HCI · Multimodal AI

Kaushika Wijerathne

Building more accessible ways to understand and interact with intelligent systems.

Position
Undergraduate Student Researcher, WVU Statler College
Degree
B.S. Computer Science, cybersecurity focus · expected May 2027
Based in
Morgantown, West Virginia · from Sri Lanka
Portrait of Kaushika Wijerathne
00About

Technology should work for more kinds of people.

I am a computer science student at West Virginia University studying how people understand, control, and experience intelligent systems. My work is shaped by severe hearing loss and by the everyday gap between technologies that technically function and technologies that are genuinely accessible.

I am especially interested in multimodal AI, spatial audio, accessible gaming, and the future of human-computer interaction as AI moves beyond standalone software and into physical, ambient, and embodied systems.

01Research

Research interests

Four questions that most of my work is a different attempt at answering.

  1. Accessible computing

    How can interfaces account for differences in hearing, perception, attention, and interaction ability from the beginning?

  2. Human-centered AI

    How can people understand what intelligent systems perceive, why they behave as they do, and when their outputs should be trusted?

  3. Multimodal and spatial AI

    How can models reason across sound, vision, space, movement, and uncertainty in dynamic environments?

  4. Embodied and ambient interaction

    As AI enters devices, homes, vehicles, wearables, and physical environments, how should people control and experience it?

02Publications

Publications & work in progress

ET-AudioBench: evaluating spatial audio reasoning in multimodal AI systems

Wijerathne, K. · advised by Dr. Anthony Sicilia · WVU SURE Research Program, 2026

Pilot built Paper in progress

ET-AudioBench is a benchmark and experimental pipeline for studying whether audio-language models can interpret spatial sound in interactive environments. The work translates an accessibility question into five reproducible task families covering direction estimation, event classification, motion and trajectory reasoning, ambiguity, calibration, and real-time control.

The pilot uses 100 clips, deterministic manifests, structured JSON outputs, specialist baselines, latency logging, and robustness conditions. Live-control scaffolding connects model performance to accessibility and competitive-game fairness for Deaf and hard-of-hearing players.

Research questions

  • Can multimodal models reliably determine where a sound originates?
  • How robust are they under noise, ambiguity, and conflicting spatial cues?
  • Can outputs support accessible interfaces or agents in real time?

Methods

  • Python evaluation pipeline
  • Stereo spatial audio
  • Audio-language model evaluation
  • Specialist baselines
  • Calibration & uncertainty
  • Interactive game environments
Structured output, one clip
{
  "direction":  "left",
  "event":      "footstep",
  "confidence": 0.82,
  "action":     "turn_left"
}
03Projects

Projects

Systems I have designed and built end to end, where the interaction problem — how a person understands, approves, and stays in control of what a machine is doing — is the hard part.

Liaison

Open-source, self-hosted AI phone agent · TypeScript Source

An AI phone agent that never hides a decision in audio.

Liaison calls customer support for you and reports back entirely in text. You describe the problem in plain language, it hands back an inspectable plan, you approve it, and it makes exactly one call — narrating the whole thing over SMS or a web thread and asking before it does anything that matters.

The design constraint is accessibility-first: nothing important ever exists only in audio. Every warning, decision, commitment, and state change reaches the person as text they can read, quote, and export — which matters most for people who find calling difficult because of hearing loss, anxiety, a speech difference, or a language barrier.

Consequence tiers

Informational
A one-line update.
Low consequence
Reply A / B / C from SMS or the web.
Sensitive
Authenticated web review only. Never over SMS.
Material
Web review plus explicit confirmation.
Prohibited
Refused. There is no path that approves it.
  • Fastify & React
  • SQLite
  • Twilio ConversationRelay
  • Deterministic policy engine
  • Human-in-the-loop AI

DELEGATE

Daily browser game · live playdelegate.com

The daily General Assembly vote game.

Predict how five countries voted on three real United Nations General Assembly resolutions. Fifteen votes, one daily score — a compact game about diplomacy, history, and reading the world.

SkillTree

In development

Details to follow.

04Writing

Writing

Long-form essays on technology, institutions, philosophy, and policy.

  1. Technology & institutions · August 2026

    The Reactor Is Not the System

    What Civilian Nuclear Power Can Learn from the U.S. Nuclear Navy

    An argument that nuclear safety is produced by institutions, training, accountability, and lifecycle stewardship — not by reactor design alone — and that civilian power should translate the Navy's reliability principles without importing military secrecy or command structures.

  2. Philosophy & lived experience · August 2026

    Grace

    The Phenomenology of Unconditional Love

    A personal essay about the distance between understanding unconditional love as a rational principle and experiencing it as a durable practice, drawing on Sri Lankan ritual, secular ethics, religious traditions, and the embodied work of making grace livable.

  3. History & public policy · August 2026

    The Brick and the Border

    Blowback, Intervention, and the Politics of Migration

    An examination of how U.S. intervention can shape the conflicts and migration later treated as isolated border problems, arguing for a politics that preserves individual responsibility while refusing to erase upstream institutional history.

05Experience

Experience

  1. Jul 2026 — present

    Undergraduate Student Researcher

    WVU Statler College

    Evaluate large multimodal AI across audio, vision, code, and text; design real-time evaluation interfaces, reproducible benchmarks, and tooling for model testing and performance analysis.

  2. May — Jul 2026

    SURE Program Researcher

    WVU Summer Undergraduate Research Experience

    Designed ET-AudioBench's Python evaluation pipeline and 100-clip pilot for spatial reasoning in audio-language models, advised by Dr. Anthony Sicilia.

  3. 2026 — present

    Cyber-Resilience Fellow

    WVU Cyber-Resilience Resource Center

    Applied cybersecurity projects supporting West Virginia organizations, coordinating with students, faculty, and project partners on practical cyber-resilience work.

  4. 2025

    Researcher and Presenter

    WVU Eberly College

    Qualitative research on hearing-impaired Sri Lankans, digital networks, representation, and accessibility. Received top undergraduate research recognition and a first-place showcase award.

  5. Aug 2025 — present

    Fitness Technician

    WVU Campus Recreation

    Maintain, inspect, and troubleshoot fitness equipment to support a safe, reliable, and accessible recreation environment.

06Beyond

Service & beyond

West Virginia Women's Alliance
Distributed menstrual products and supplies at community events.
FeederWatch
Contributed seasonal bird observations to community science.
U.S. Geological Survey
Contributed observations to public geological and environmental science.

Top-ranked across several competitive esports titles — Valorant, Marvel Rivals, Apex Legends. These games make spatial sound consequential, which keeps the accessibility research grounded in real interaction.

07Contact

Let's build more accessible systems.

I am interested in research collaborations, PhD opportunities, accessibility work, multimodal AI, HCI, and conversations about how intelligent systems should interact with people.

kaywijerathne@gmail.com