Anvil Sim
Engineering decisions need receipts.
Anvil turns simulation output into a review surface: the model, solver result, benchmark delta, and decision context should be visible together.
Anvil Sim + Rove Dexterity Lab + Enterprise OS
Building the stack for physical AI.
From engineering simulation to robot dexterity: Anvil Sim is the simulation wedge, Rove Dexterity Lab is the hardware wedge, and Enterprise OS is the execution layer behind the lab.
The goal is AI-native engineering infrastructure that moves from model to measurement to real-world action.
Model
Simulation
Understand the physical system before touching hardware.
Measure
Receipts
Make claims inspectable with benchmarks, deltas, and replayable outputs.
Act
Dexterity
Close the loop with robot tasks that can be repeated and improved.
Flagship labs
Simulation first. Dexterity next. Physical AI over time.
Lightbulb starts where physical AI has the highest leverage: tools that make physical systems inspectable, then hardware tasks that make intelligence measurable in the real world.
reach envelopebench prototype
Rove Dexterity Lab
Dexterity begins with measurable tasks.
- Manipulation bench before full-body robotics
- Gripper and claw experiments for small-object handling
- Task ladder for grasping, sorting, pressing, and recovery
- Simulation used where it shortens the path to real hardware
Proof wall
The stack only matters if the proof is visible.
Anvil carries the strongest current evidence: real solver footage and reviewable receipts. Rove stays honest as a hardware track until the physical bench earns its own footage.
Simulation wedge
Anvil decision loop
Model, result, comparison, and receipt in one reviewable surface.
Physics receipts
Modal / FRF inspection
Motion makes solver output inspectable instead of hiding it in tables.
Physics library
Thermal transient
A smaller physics track that expands the stack beyond structural demos.
Build queue
Model to measurement to action.
Now
Turn Anvil into decision infrastructure
Move from proof reels into reusable review bundles: model, solver output, delta, receipt, and engineering decision.
Next
Build the Rove dexterity bench
Start with physical tasks that can be repeated and measured: reach, grasp, sort, press, recover, and report what happened.
Layer
Use Enterprise OS as the execution layer
Keep research, specs, agents, evaluation, and release cadence connected so the lab compounds instead of drifting.
Internal engine
Enterprise OS is the execution layer.
It handles research intake, memory, evaluation, agent loops, and shipping discipline. The public story is Anvil and Rove. The internal edge is turning research into shipped prototypes faster than a solo founder normally should.
Founder
Built by a founder turning AI into physical execution.
Ray Qin
Founder & CEO
Dual ME and CS background. Senior software engineer building the stack for physical AI: Anvil Sim for engineering simulation, Rove Dexterity Lab for robot manipulation, and Enterprise OS as the private execution layer behind the lab.
Lightbulb exists to make physical AI more concrete: fewer slogans, more measurements, clearer receipts, and demos that move from model to real-world action.
Contact
Talk if you are building physical AI, simulation, robotics, or useful hardware.
Looking for collaborators in engineering simulation, robot manipulation, useful hardware-software prototypes, and evaluation loops that turn intelligence into action.