AI Researcher &
Machine Learning Engineer

ASPIRE — Real-time accessible slide presentations

macOS app for blind and low-vision audiences at live talks. Captures screen and system audio (ScreenCaptureKit), detects slide changes with perceptual hashing and OCR dedup, writes a short description with a multimodal LLM, speaks it with on-device TTS at the next natural pause, and reclaims time by trimming pauses and filler words. Streams to listeners' phones over WebRTC or HLS. In a pilot, content recall rose from 16.7% to 91.7%.

Gap Sense — Street-crossing practice for blind travelers

An orientation-and-mobility training tool that teaches how long a crossing takes as a felt duration: the student taps at each waypoint and hears chimes or buzzes where their timing was in or out of tolerance, never a number of seconds. Ships as a web app, an iOS build (TestFlight) with Taptic Engine haptics, and an Android APK.

EmoBeat — Temporal-aware multi-label mental-health screening

Built EmoCorpus, a 30,192-post Reddit dataset annotated with seven DSM-5-aligned conditions, a temporal-arc taxonomy, and a recovery schema via an LLM-assisted pipeline (98.5% label agreement). Designed a hierarchical sliding-window transformer with tense-aware attention pooling and a recovery-penalty loss; benchmarked 11 architectures across 5 model families (83.2% Macro F1, 94.2% ROC-AUC) and showed that every baseline suffers "trajectory blindness" on recovery narratives.

TruthfulQA truthfulness classifier

Repurposed TruthfulQA into a balanced 1,580-example supervised dataset and fine-tuned DistilBERT to 89.5% test accuracy and 0.892 F1, 25.5 points above a word-embedding baseline. Built an API-free scorer that assigns continuous truthfulness probabilities to LLM outputs; applied it to 790 GPT-4o-mini answers and validated against 100 human-annotated answers (Cohen's κ = 0.86, Spearman ρ = 0.78).

Interactive Computing Lab website

Built and maintain the public website for the Interactive Computing Lab at Stevens (Dr. Jonggi Hong): team, publications with filters, courses, and openings, plus a maintainer guide so lab members can update it without tooling.

This website

Hand-built static site with an interactive AI Playground (neural network forward/backprop, embedding space, self-attention heatmap, optimizer races on loss surfaces), a live computer-vision kernel demo on the profile photo, and a grounded chatbot served by a Gemini model behind a Cloudflare Worker.

ACM ASSETS 2026 · Poster · To appear

ASPIRE: Accessible Slide Presentation via Intelligent Real-time Editing

Azizul Haque, Jonggi Hong. A fully automated, real-time realization of delay-buffered editing for live talks. Across four full-length talks and a pilot with four blind and low-vision participants, content-recall accuracy rose from 16.7% to 91.7%.

ACM DIS 2026 · Full paper · 🏅 Best Paper Honorable Mention

Enhancing Slide Presentation Accessibility for Blind and Low-Vision Audiences Through Delay-Buffered Editing

Azizul Haque, Jonggi Hong. Proceedings of the 2026 ACM Designing Interactive Systems Conference. A delay-buffered editing approach that trims redundant speech and inserts concise slide descriptions at transition points within a five-second buffer, with an automated pipeline evaluated with blind and low-vision participants.

ACM ASSETS 2025 · Poster

Accessible Slide Presentation via Intelligent Real-time Editing: Insights from a Wizard-of-Oz Study

Azizul Haque, Jonggi Hong. Proceedings of the 27th International ACM SIGACCESS Conference on Computers and Accessibility. A Wizard-of-Oz study with blind participants showing that a short temporal buffer makes room for slide descriptions and improves transition clarity and comprehension.