A low-code local genAI-powered programming interface that explores the use of LLMs as natural-programming language compiler/translator.

• locally-run code LLM (Qwen2.5-Coder)
• generated code is evaluated in-browser with Wasm
• intermediate pseudocode prioritizes understanding
• low-code node-based programming interface

A prompt-less deterministic image genAI which takes math expressions of images as input, based on prior CV research.

• simple drag-and-drop no-code interface
• job queue dispatches API calls to GPU backend
• uses server-sent events (SSE) for real-time frontend updates

A vectorDB, RAG and genAI-powered note-taking progressive web app (PWA) for writers to understand themselves. In closed beta.

• auto-identifies connections across notes
• generates tailored consolidation and recommendation using RAG¹
• I use it as my daily driver

MIT 6.5820 (Networks) research exploring delay and network probing tradeoffs for real-time video.

• used existing BBR implementation in Linux kernel
• identified optimal probing strategy given constraints

A scheduling feature update enabling drag-and-drop (DnD) interactions and task assignment.

• edit element dates on calendar using DnD
• view task assignees

A one-click genAI & RAG feature that crafts a presentation using project contextual metadata.

• understands project context and available assets
• produces a pitch-ready Google Slide or .PPTX
• provides options for specified or freeform structure

A system that ingests multiple file types to construct project-specific contextual metadata for downstream genAI and RAG tasks.

• ingests wide range of file and link types into common metadata format
• auto-ingestion of project asset updates
• high throughput via serverless architecture
• fast metadata production by parallelizing API calls

An alternate YouTube watcher that prioritizes user attention while adding useful features.

• prioritizes user attention by reducing distraction
• user-customizable settings (e.g. time limits)
• users can curate channels by selecting creators

MIT 6.8300 (CV) research exploring diffusion model latent space navigation using images as input.

• inverted images into CLIP embeddings for arithmetic computation to generate final image
• found formulaic expressions that extract semantics and style
• potential for non-technical users to probe the model (e.g. for bias) – see PixMath

Research exploring projection-based augmented reality (AR) for explainable robot vision (CV).

• combines projection-mapping and real-time object detection for model transparency

MIT 6.8610 (NLP) research exploring human-interpretable concepts inside contextual word embeddings.

• used a simple linear classifier to identify embedding axes related to specific concept
• found that concepts are entangled across axes
• identified axes with greater impact on classification

A location-based AR web app that digitally preserves places that are lost through urban renewal.

• web-based AR, does not require app download
• location-based archival image retrieval
• a DesignSingapore Good Design Research grant project

A real-time data viz app that automates business insights (BI) from DocMed's e-commerce platform.

• instant access and visibility of BI
• optimized for low-compute smart TVs

A genAI p5.js editor/REPL for coding beginners learning to program visually.

• immediately visualizes generated output

A simple web utility that keeps time across timezones, useful for globe-spanning teams.

• show the current time or wind it forward or backward