This view needs WebGL.
click a room
I'm Tarunesh. I work on AI interpretability for scientific discovery.
17, Prosper, Texas. Research engineer at Aquin Labs. Founder of Pathway Counselor, co-founder of Young Tech Pioneers and Modularity. Research at UMBC and UC Santa Cruz, published at NeurIPS, ICML and IEEE BigData.
what I believe
The world isn't short on intelligence, human or otherwise.
It's short on people who turn what's already known into something others can actually use.
Foundation models are monolithic: improving one skill means retraining and re-testing everything. Modularity builds models whose capabilities can be added, removed or upgraded on their own, so a robot that fails at a new task fixes that one piece instead of starting over.
Most high schoolers never get real college counseling. An AI assistant that answers from a district's own documents and a large scholarship database, so the advice is specific and checkable. 500+ students.
Aquin builds language-model interpretability tooling for other labs. I'm a research engineer there, shipping tools to biotech and research groups.
NVIDIA Inception
Selected as a technical builder and flown out to SF, with $25K+ in AI and cloud credits. Picked from thousands for the VibeCon SF hackathon and placed top 10. The afterparties were good too.
YC Startup School
Mechanistic interpretability, AI for science, hallucination, trustworthy AI. One question from four sides: what did the model learn, and how would we check?
Can a model tell whether a patient will recover from a coma, and show its reasoning? COMA reads brain activity through concepts it found on its own, and beat the standard clinical guidelines at hospitals it had never seen.
UMBC KAI² Lab, first author
A language model never trained on a rare disease still knows something about it. I read symptom profiles out of a frozen model's internals, and they beat a purpose-built medical model's.
Accepted at NeurIPS 2026, first author
Does a model see a scene as parts, or as one blur? Edit one object and check whether its explanation still holds. A benchmark I co-led at UC Santa Cruz.
ICML 2026, co-author
A model can recognize a concept without that concept driving its answer. I showed this for "threat" and "blame" in a small Llama model at Aquin, and wrote down how to tell the two apart.
Accepted at NeurIPS 2026, first author
A neural network trained on thousands of real human choices predicts them better than the leading theory in psychology.
Accepted at IEEE BigData 2026, sole author
A nonprofit I co-founded in ninth grade so students could try running a real business: chapters, fairs, a marketplace to sell on, and direct funding, including five businesses in Kenya.
Featured on FOX 4 News
I lead Redacted, a CTF team: cryptography, reverse engineering, web exploitation. 9th in the US at picoCTF, 2nd nationally at the US Cyber Challenge.
picoCTF, Carnegie Mellon