Dillon Mehta

CueBench

What I did. Co-founded a YC S26 company building reinforcement-learning environments and deterministic verifiers for AI models. Our work includes PerfGym and Quantum Hamiltonian Bench.

What I learned. A technically sound environment is only a start; the harder problem is building it reliably for others to use at scale.

Maze robot

What I did. Designed and soldered the PCB, built the shell, and wrote C++ motion control using encoder and IMU feedback for timed maze navigation.

What I learned. Mechanical variation, sensing, and control code have to be tuned as one system. Three hardware iterations earned seven medals.

VEX climbing system

What I did. Designed the hang, drivetrain-to-winch power sharing, and odometry hardware. At Worlds, the robot reached the third rung in under 12 seconds and scored a ring.

What I learned. Mechanism design succeeds when power, traction, and packaging work together. My teammates wrote the VEX software.

Medical AI research

What I did. Built the base multi-agent diagnostic simulation and directed ablations testing how outside advice and bias affect decisions.

What I learned. Final accuracy hides how a diagnosis changes. The timing and repetition of an intervention can matter as much as its content.

Matrix multiplication unit

Learning phase. I am learning chip design and design verification while working toward my own matrix multiplication unit.

I have not finished the build. Come back to see what I make and what the verification work teaches me.