You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
$45,000 left to reach the funding goal.
spent twenty years building low-latency trading engines on Wall Street. In that world, rule number one was that software running on the same CPU as the process it monitors will always get bypassed when things go wrong.
Yet right now, every AI safety guardrail is pure software. Docker sandboxes, prompt filters, supervisor agents—they all share the host CPU and memory bus with the model. If an agent loops uncontrollably or exploits a zero-day, the sandbox still has to ask the operating system for cycles to terminate it. That won't hold up.
I built a physical circuit breaker on an FPGA card instead.
It’s called sre_watchdog_shield.sv. It sits between the host controller and the accelerator. If an agent loops, ignores a kill signal, or attempts unauthorized network traffic, a hardware counter trips and drops the physical bus lines in 50 milliseconds. Software can't touch it because the gates sit outside the CPU memory map.
Alongside that, I wrote cred_gate_ctrl.sv to token-gate on-die SRAM every clock cycle so adversarial prompts can't overwrite weights, and omega_mollifier_core.sv, which runs in 264 nanoseconds on AMD Alveo FPGAs without touching off-chip DRAM.
The verification logs and test benches are already public on CERN Zenodo under DOIs 10.5281/zenodo.24071603 and 10.5281/zenodo.24074626.
I've paid for all my lab equipment and FPGA cards myself so far. I'm asking for $15k–$45k to put this into an open repo so safety researchers can red-team it.
The money goes straight to renting AMD Alveo cloud instances for adversarial fuzz testing, open-sourcing the RTL and AXI4 wrappers on GitHub under an open license, and paying bounties to anyone who can break the 50ms cutoff.
Happy to answer questions or hop on a call with anyone interested in hardware-layer safety.