AI research fellow at Purdue University, studying artificial intelligence. I work on what machine learning models learn and what they give away: secrets that leak out of LLMs, agent swarms that translate Sanskrit, music transcription and anomalies in GPS signals.
Scroll, and the cloud reshapes into each project.
Fig. 1. A neural network mid forward pass. Each column is a layer of neurons, and brighter lines are stronger weights; amber traces the neurons that fire as one input flows through to a prediction. Every figure below is this same cloud of points, reshaped.
NSF ACTION AI Institute
UC Santa Barbara SecLab, with Dr. Yigitcan Kaya
Can an LLM leak a secret without ever saying it? In my first-author paper, accepted at the NeurIPS 2026 InfPriv workshop, a white-box likelihood attack identifies a hidden PIN from one strictly filtered list of numbers, and recovers 48% of held-out PINs from repeated outputs. I'm continuing this work on token entanglement and subliminal learning toward a follow-up paper.
Read more about ACTIONFig. 2. A causal attention mask over 48 tokens. The amber cells rising off the plane are attending to a hidden system prompt.
Berkeley AI Research
Dharmamitra, with Prof. Kurt Keutzer
Machine learning that translates underserved Asian languages into English. I built MITRA-Swarm, a multi-agent architecture that injects explicit morphophonemic and discourse structure into foundation models for classical Sanskrit translation. It beats compute-matched self-refinement baselines under an independent two-judge evaluation, and my lead-author paper was accepted at the NeurIPS 2026 LP4FM workshop.
Read more about BAIRFig. 3. संस्कृतम्, Sanskrit, sampled from a Devanagari typeface. The amber stroke is the shirorekha, the headline that joins the letters of a word.
AI4Musicians Lab
Purdue University
Automatic music transcription: turning recordings into the notes that were played. I lead a research team of more than 50 members, organized the IEEE-sponsored 2025 AMT Challenge, and co-authored its results paper at the NeurIPS 2025 AI for Music workshop, benchmarking transcription models from Google, Spotify and ByteDance.
Read more about AI4MusiciansFig. 4. A transcribed piano roll, pitch against time. The amber notes are the bass line, separated out as its own instrument.
MusicRAI
University of the Arts London, with Prof. Bryan-Kinns
Responsible AI for music generation, part of an effort to reduce bias in AI music tools. I engineered and trained music generation models, then optimized and packaged them so the public could use them.
Read more about MusicRAIFig. 5. A spectrogram of a generated tone as a surface over time. The amber ridge is the fundamental; the others are its harmonics.
Governance & Responsible AI Lab
Purdue University, with Prof. Daniel Schiff
Responsible AI research on how AI is changing the workforce. I help build high-quality datasets of scholarly knowledge about AI's workforce impact, coding research papers against a structured codebook so findings can be compared across the literature.
Read more about GRAILFig. 6. A citation graph of clustered papers. The review at the centre draws amber threads to every cluster.
National Space Intelligence Center
The Data Mine of the Rockies, with NSIC and KBR
GPS is the most monitored signal in the world, and one of the most complex systems the Space Force runs. I trained a machine learning model that detects signal-to-noise anomalies in time-series GNSS data, for use by the U.S. Space Force.
Read more about NSICFig. 7. The six GPS orbital planes, inclined at 55°. The amber satellite has drifted off its track.
Autonomous Robotics Club
Purdue University
Optical music recognition: teaching a computer to read printed sheet music. I worked on homr, which uses a two-stage pipeline: segmentation for structural analysis, followed by semantic symbol recognition with transformer models.
Read more about ARCFig. 8. Two lines of printed music. Segmentation finds each symbol; the amber boxes, one layer forward, are what the transformer reads as notes.
Publications
In preparation
- Follow-up on secret leakage and subliminal learning in LLMs ACTION
- Follow-up on agent swarms for Sanskrit translation BAIR
- Automatic music transcription AI4Musicians
- AI's impact on the workforce: a systematic review GRAIL
Honors
- CRA Outstanding Undergraduate Researcher Award, honorable mention 2025–26
- Dean’s List and Semester Honors 2024, 2025, 2026
Fig. 9. Back to Gaussian noise, where every figure started.