/ PROJECTS

Projects

Robots, brains, companies, and lately some quantum. What I'm most focused on, then everything else.

RoboticsUnder review2026

GAIN

Grounded Active Interactive Navigation

A robot that works out it has to open the door or call the elevator, and then does it.

90.8%door trials completed, in sim
86.7%elevator trials in sim, vs 41.7% for a plain VLM
85%long missions chaining both, in sim
How it works

A foundation model decides what has to change in the environment; separate grounding and control modules decide where and how to act. Tested in Isaac Sim (ROS 2 / Nav2), then on a Ranger Mini 3 with an xArm6 that rides elevators between floors in a campus building. First author, with Dr. Minghan Wei at FAU's SIMA Lab. Submitted to IEEE ICRA 2027.

Code on GitHub ↗Under double-blind review, so the paper isn't posted yet.
INKY public demo with sample data
AIPrivate beta2026

INKY

AI study tutor for high schoolers

A tutor engineered so it won't just give the answer, with separate models checking both the student's message and its own reply.

774commits on the development branch
~130klines of TypeScript and SQL
46edge functions
How it's engineered →Public demo ↗
First page of the PhysMent paper
AIPublishedJul 2026

PhysMent

AI for MATH Workshop, ICML 2026 · PMLR 306

Static benchmarks only test what a model has read. PhysMent makes language models run experiments in a live physics simulator to find the answer.

105classical mechanics scenes
26simulation tools the models call
25–67%accuracy across seven models
The finding

Models apply forces, query object states, advance time and change the scene in MuJoCo before answering, and the physics engine verifies the result. They reached up to 80% on qualitative single-concept tasks, but most fell below 30% on the hardest. The failures were mostly about procedure: premature answers, inefficient exploration, and not grounding in what the simulator showed. Co-author, 5th of 9, through Algoverse AI Research.

Read the paper →ICML page ↗
Anik and judges holding a Diamond Challenge check made out to AuthAI
VenturesFounder & CEO2025–

AuthAI

Prior authorization for orthopedic practices

My dad is a physician, and I watched him stay up late fighting prior-authorization forms. AuthAI drafts clinical notes and checks them against insurers' prior-auth rules.

1stDiamond Challenge, New York regional; international finalist
State 1stDECA Independent Business Plan
Runner-upFAU Business Pitch, Student Track
Pitch coverage ↗Podcast ↗
Anik beside the FAU autonomous golf cart
RoboticsActive2026

Self-driving golf cart

2018 GEM e4, with Dr. Hahn at FAU

Turning an electric cart into a vehicle that drives itself, built from the control stack up.

ROScustom control stack
Full stackwiring, firmware, perception
Firststeering and braking under computer control
Case study →
Quantum3rd placeOct 2026

aperion

Quantum-assisted ultrasound brain imaging

Moving the slowest step of ultrasound stroke imaging onto a quantum computer.

0.58%error vs the exact answer, ideal circuit
78%less pressure error on IBM hardware
135qubits in the 3D circuit
How it works

One ultrasound brain image takes about 32 hours on 128 computers, and over 95% of that is simulating the sound wave. We rewrote the wave equation as a Schrödinger equation and built that step as Qiskit circuits, then ran it on IBM hardware with gate twirling and a physics-aware error filter. Classical still wins at today's sizes; the advantage needs error-corrected machines. 3rd place at the IBM Quantum Hackathon (Florida Qiskit Fall Fest), with Aaron Loh and Charan Jarugula.

Code on GitHub ↗

Everything else 12 projects