Xué huì le! Tap a word to hear it.
Your Earth won’t survive this. A rift to the other Earth has torn open at the end of the street below.
Run for it and live on the other Earth, or stay and go down with yours.
Thanos’s snap threw you across the galaxy. You’re floating in the wreckage near Titan.
Teleport home and the snap stays. Or find all six Infinity Stones and snap everyone back yourself.
Type a sentence. It will be split into tokens, and you can watch attention jump between them.
Click a surface to place a block. Drag to turn the view. Right-click or turn on Erase to remove blocks.
If the site doesn't load here, use "Open in a new tab".
Which radical is hiding in this character?
I'm a Grade 12 IB Diploma student working where AI, machine learning and language meet. I'm fascinated by how intelligent systems process human language, and how to make them fairer, safer and more useful. I've published research, built apps for farmers and Mandarin learners, and interned as a machine learning engineer.
Finishing the IB Diploma (Math AA HL, Physics HL, Chemistry HL, Economics, English Language & Literature, and Mandarin), graduating in June 2027.
Seven years of Chinese and counting. I'm working at roughly HSK 4 to 5 level, with a 2,500+ word vocabulary, which is why language keeps showing up in my research and projects.
Hip-hop on repeat, a growing vinyl shelf, a Batman figure collection, and far too many hours in Minecraft (Java Edition, obviously).
Fairness and interpretability in AI.
Can synthetic data safely stand in for real data in loan approval models? I compared SMOTE, CTGAN, TVAE and Gaussian Copula on the German Credit dataset and created a Dependency Preservation Score to measure how well each kept the relationships between credit variables intact.
The finding: every synthetic method distorted those relationships and made fairness less stable, so for high-stakes decisions like lending, original data with fairness-aware preprocessing is the safer choice. The work earned a Gold CREST Award from the British Science Association.
Read the paper PDFTech built for real people.
A multilingual Android app for Indian farmers earning under ₹10,000 a month. It helps them find nearby markets, government schemes and credit information in their own language, with a chatbot for quick questions.
I validated it with NABARD and three Farmer Producer Organisations in Karnataka, raised ₹1,20,000 to build it, and early results suggest around a 10% rise in income. AgriBridge is now available to farmers in every state in India. I presented it at the 64th UN Commission for Social Development in New York as a 1M1B Changemaker.
A full-stack web app for mastering radicals, the building blocks of Chinese characters, with interactive flashcards, AI-powered hints, timed challenges and animated stroke-order guides. The little scroll game in this room is a taste of it.
Try MandalinkVolo processes health insurance claims, where data is full of sensitive personal and health information. I built document classification to check claim completeness, Aadhaar number extraction for CKYC, and Aadhaar masking to support privacy and DPDP Act compliance, cutting manual work and errors.
What I've had on repeat this year, live from Last.fm.
Everything in the crate.
What I think about what I'm listening to.
News from Mandalink and this site, live from the Mandalink board.
Loading announcements…
Awards, activities and a few things I'm proud of.
Want to talk AI, language, music, or anything else?
contact@rayanupadhyay.comIn the pool since Grade 3.
I've been on my school swim team since Grade 3, with medals in backstroke, individual medley and relay.
A few moments.
Two doors out of here: the moon gate to Mandalink world, and the hospital doors to Meridian General. And here's what's on the table.
New research on how AI models process emotion. Details after publication.
Something new. Check back soon.
Grade 12 IB Diploma student in Bangalore working where AI, machine learning and language meet. Published researcher, ML intern, and builder of apps for farmers and Mandarin learners. Graduating June 2027.
Compared SMOTE, CTGAN, TVAE and Gaussian Copula on the German Credit dataset and created a Dependency Preservation Score to measure how well each kept the relationships between credit variables intact. Every synthetic method distorted those relationships and made fairness less stable, so for high-stakes lending decisions, original data with fairness-aware preprocessing is the safer choice.
Health insurance claims: document classification to check claim completeness, Aadhaar number extraction for CKYC, and Aadhaar masking for privacy and DPDP Act compliance.
Multilingual Android app helping farmers earning under ₹10,000 a month find markets, schemes and credit information. Validated with NABARD and three FPOs in Karnataka, raised ₹1,20,000, early results suggest around a 10% rise in income. Presented at the 64th UN Commission for Social Development in New York as a 1M1B Changemaker.
Full-stack web app for learning Chinese radicals: flashcards, AI hints, timed challenges and stroke-order guides. Python, Streamlit, Groq, edge-tts.
mandalink.orgThe whole site is one HTML file with no build step. The 3D room, Mandalink island, the rooftop and Bangalore are drawn with Three.js, and every model (the furniture, the creatures, the weapons, the city) is built from simple shapes in code rather than downloaded. Sound effects are generated live with the Web Audio API, progress is saved in your browser, and other visitors show up as ghosts through a small public message relay. There is a simple version for screen readers and keyboards, a reduce-effects setting for slower devices, and I test it with a headless browser.