Trustworthy AI · retrieval · physiological ML · language technology · weirdly useful experiments
Python · TypeScript · Next.js · Supabase · Computer Vision · Research Engineering
▼ click any mode below to open its panel · click again to close
🧪 R&D MODE — research, hypotheses, benchmarks, failure analysis · click to open
Research first. Metrics second. Failure cases always.
I’m interested in the part of AI research that asks not only “does it work?” but also where it fails, whether the evaluation is honest, and what evidence would actually falsify the claim.
|
Falsification-driven reasoning for LLM research Competing hypotheses, explicit predictions, executable tests, failure conditions, and provenance-aware evaluation.
|
Premature-closure research system Controlled experiments around reasoning behavior, blind evaluation, reproducibility, and structured benchmark runs.
|
|
Evidence-aware AI systems Explores how generated answers can remain tied to inspectable evidence instead of treating fluency as proof.
|
Search for things in the physical world Vision + spatial-memory experiments for finding previously seen objects while reducing false alarms and wrong-place retrieval.
|
|
Subject-conditional conformal prediction for biosignal ML Focuses on per-subject reliability rather than hiding difficult cases inside population averages.
|
Personalized calibration for wearable stress detection Studies subject-specific baselines, LOSO evaluation, selective prediction, and when personalization does not help.
|
- When does a better average metric hide a worse failure mode?
- Can uncertainty guarantees remain reliable for individual subjects?
- How should retrieval systems behave when the “closest” context is still wrong?
- Can AI systems expose limitations and provenance instead of pretending every output is equally trustworthy?
⚙️ DEV MODE — products, deployed systems, interfaces, shipping · click to open
Ideas are cool. Shipping them is cooler.
This is where research leaves the notebook and has to survive auth, databases, interfaces, edge cases, deployment, and actual users.
|
GKS discovery + preparation platform A full-stack system for applicant discovery, scholarship information, interview preparation, privacy-aware profiles, and community workflows.
|
Korean-learning companion A language-learning product built around guided practice, conversational interaction, and a more human-feeling learning experience.
|
|
Multilingual browser reading assistant Word-level reading support for Korean, Japanese, Chinese and other languages through a Chromium extension.
|
Context-first Korean learning Video-based listening and context practice with level-based content, guest flows, personalization, and creator/community features.
|
🌙 2:00 AM MODE — odd ideas, tiny escapes, and experiments that got out of hand · click to open
It's 2 AM. The idea is ridiculous. The repo already exists.
A Minecraft sidekick with a memory. A club for doing the bare minimum. A way to remember who you met and why it mattered.
playful on the outside · intentionally built on the inside
|
An AI Minecraft companion that remembers your world. Not just current stage: V0 scaffold · autonomous errands are next
|
For when your brain has 2% battery and 37 tabs open. Tiny distractions, low-effort comfort, the Regret Button, character-development lore, and quiet company. No motivational speeches required. vibe: no fixing · just surviving
|
|
“Wait, who was that—and why should I talk to them?” An opt-in event networking utility for remembering people, finding genuine overlap, planning a first conversation, and knowing when to follow up. superpower: turning awkward introductions into context
|
Perform an autopsy on a conversation. A playful conversation-analysis experiment with a forensic interface and deliberately dramatic personality. status: somehow works
|
|
You've been online too long. Prove you still exist. Instead of proving you're not a bot, this tiny anti-doomscroll experiment asks you to prove you're still acting human. status: alive
|
If it keeps coming back to mind, build a tiny version. Some ideas become tools. Some become experiments. Some should probably have stayed in the notes app.
|
I’m Susan, an incoming undergraduate student from India preparing to study Artificial Intelligence / Computer Science.
I’m especially interested in how AI systems fail, how uncertainty should be communicated, how retrieval can go wrong, and how research ideas can be turned into usable systems.
I’m still early in my academic journey, so I care more about developing good research habits than presenting myself as an expert.
Curious enough to build · careful enough to question the result