~/aarushgoradia.com princeton, nj
aarush goradia

aarush goradia

$whoami

i work where models meet silicon: computer architecture, inference systems, and the verification that keeps chips honest.

$cat about.txt

senior in ece at princeton, minoring in cs + english.
co-founder @ murphy labs, where we break robots before the real world does.

~/now// what i'm working on

senior thesis · trillion-param moe inference on commodity hardware

the whole model sits in nvme nand. active experts get staged through dram into gpu memory. my bandwidth model says sequential decode is stuck around 1.5 tok/s on our hardware, so i'm modifying llama.cpp to get past 10 tok/s. advised by david wentzlaff.

× 3 nodes · expert cache · prefetch · compute/io overlap

veri2

co-first author. rtl→sva training data filtered by jaspergold proofs, then qwen2.5-coder-7b fine-tuned on it. 50.2% formal yield vs gpt-5's 41.4%, with ~200x fewer params. planning to put it on arxiv.

murphy labs

co-founder. breaking robots before the real world does.

~/work

2026.06–08

qualcomm · digital verification intern

formal (jaspergold + sva) and uvm for noc power management. built multi-agent llm debug tools three teams adopted, weeks → hours. return offer.

2025.06–07

advanced micro foundry · tech dev intern, singapore

automated gds fab checks + drc for tape-out. c++ rewrite: 15–20x faster.

2024.09–now

princeton rover · electrical

servo power and bus bars that move a lot of battery current, built to be safe to work next to.

2024.05–07

ai seer · ai engineering intern

langchain fact-checking agents (85% → 92%) and realtime whisper transcription. filed a us patent.

~/research// & builds

2024–25

branch prediction on a 5-stage risc-v core

bimodal, gshare and tage-lite. gshare cut nested-loop mispredicts from 90.6% to 3.8%, but decode-stage redirects ate most of the ipc win.

2025

slam rc car

ros2 + cartographer on a raspberry pi 5, with lidar/imu fusion and apriltag waypoint navigation.

2024–25

rf transformer synthesis · sengupta lab

automated layout sims in virtuoso/emx and trained a net to map lumped parameters back to geometry, 90% accurate.

~/film// letterboxd top 4

~/misc

reading
the magic mountain, thomas mann
notes
llm-wiki, my research wiki for comp arch + ml systems papers, kept up with claude code
also
pc building, robotics, strategy games, literature