Screenshot 2024-08-14 at 10.09.02-02.png

Twenty years early. Three shifts. One job.

Top-10% national sales award at ADP. The youngest and only woman on SAP's enterprise software team. I cold-DM'd my CTO on Twitter, three months later we had a pipeline running on a real robotic arm.

WHAT I'M BUILDING NOW

Lisa Cheng · Cofounder & CEO, Arithmancy

ABOUT

Who is Lisa Cheng?

Twenty years, three shifts, one job.

I've spent twenty years being early to technologies before they were investable. Crypto in 2013. AI in 2024. Robotics now. Each time I've played the same role: the person in the room where the decision actually gets made.

I learned to be that person the hard way. In 2001 I was building MySpace pages and teaching myself HTML. In 2007 I thought I wanted to be a lawyer, so I wrote the LSAT on a whim. In 2009 I got into software sales, I won a top-10% national sales award at ADP, then was recruited to SAP as the youngest, and only, woman on the enterprise software team. That's where I learned the thing that now runs Arithmancy: revenue happens in a room, with a person, who has to trust you.

In 2013 I read the Bitcoin whitepaper and never looked back. I've been early to every shift that followed.

I learned governance the hard way too. Arithmancy is built the opposite way, measured, audited, honest from first principles.

In 2024 I joined an AI company. In 2026 I cofounded Arithmancy, and built the instrument that catches what every simulator misses.

The three shifts, on one line each:

Crypto 2013. I read the Bitcoin whitepaper, found a bounty on bitcointalk, completed it, and was hired by the first company to do an ICO.

AI 2024. I joined an AI company as a blockchain architect and subject-matter expert.

Robotics 2026. I own go-to-market, partnerships, and the demand test that decides what we sell.

Arithmancy measures the gap between a robot's simulation and its hardware.

Every robotics team trusts a simulation it has never audited. The simulator tells the policy exactly where every object is. The real robot has to act on whatever its sensors can reconstruct, and that's never perfect. So the policy clears in simulation and misses on the arm, weeks later, in front of a customer.

We built the instrument that measures where a simulation diverges from the hardware it claims to represent, and names the defect before it costs a robot. On its first run it caught two.

One was a collision-mask bug. A policy recorded 92 out of 100 successful grasps, then collapsed to zero the moment the simulator's actual settings were honored. The other was a grasp that holds in classic MuJoCo but ejects the object the moment you migrate to MuJoCo-Warp or Newton — the commands drive the pads 9.8 millimeters into a brick that's only 15.8 millimeters thick. The simulator never noticed. The robot would.

We never declare which engine is right. That's not ours to claim, and it's not something we sell. We find disagreement, not truth.

I found my cofounder, Mustafa Khan, by cold-DMing him on Twitter. Three months later we had a pipeline running on a real robotic arm.