Intelligence doesn’t stop at software.

We're a small team of computer scientists bringing hardware reasoning to everyone. Led by Prof. Mark Santolucito, our team comes from a Barnard/Columbia research lab. Our platform provides cloud silicon, open-source toolchains, and training environments where AI learns to design computer chips that are both correct and optimal.

Silicon Reasoning on the Cloud

Prototype digital circuits on physical silicon in a few lines of code with our cloud FPGA cluster, Python SDK, and CLI.

Built on the open hardware stack — Yosys, nextpnr, LiteX, and the Lattice ECP5-85F.

client_sdk.py

import manhattan_reasoning_gym as mrg

class Regs(mrg.RegisterMap):
    DATA_IN  = 0x0004
    DATA_OUT = 0x0008

app = mrg.cloud.App(
    "my_design",
    design="design.py",
    registers=Regs,
)

@app.local_entrypoint()
def main():
    with app:  # programs the FPGA, releases on exit
        app.write(Regs.DATA_IN, 0xDEADBEEF)
        print(hex(app.read(Regs.DATA_OUT)))

terminal

sandbox.py

import manhattan_reasoning_gym as mrg

sandbox = mrg.Sandbox(
    files=["examples/design.py", "examples/agent.py"],
)
result = sandbox.run("agent.py")  # no network, no keys

02 · Sandbox

Train and benchmark hardware agents at the transistor level.

Isolated sandboxes that make it easy to design, experiment, and verify on real silicon in a matter of minutes. We’re currently working on an RLVR environment designed by formal methods experts.

Local Sandboxing
Verifiable Reward Signals
Real Silicon in the Cloud
Benchmarks in the Works

03 · Beta Access

What Will You Build?

Dive into real silicon with two commands.

install

$ pip install manhattan-reasoning-gym
$ docker pull ghcr.io/manhattanreasoning/mrg-sandbox:latest

Request private beta access

We’re looking for users to test our infrastructure — tell us a little about what you want to run on it.