A 50ms hardware bus interrupt and memory-isolation controller on FPGA to physically stop rogue AI tool-use and model exfiltration.
I worked 20+ years in quant trading infrastructure (JPMorgan Private Bank, Raymond James, LPL, Ally) before starting Mavericks Alpha Signal LLC in Charlotte, NC in August 2026. In low-latency trading, any risk check running on the same OS and CPU as the trading process fails when the host hangs or gets compromised.
Right now AI agent sandboxes, prompt filters, and eval monitors all run in software on the same Linux kernel and shared DRAM bus as the model. If an agent enters a runaway tool-call loop or escalates root privileges, software tripwires share the same fault domain.
In my Charlotte lab I built a hardware cutoff on AMD Alveo FPGAs instead:
sre_watchdog_shield.sv: physical circuit breaker on the 128-bit AXI4-Stream bus. Drops physical I/O lines in 50.0 ms outside the host CPU memory map.
cred_gate_ctrl.sv: 64-bit cycle-by-cycle hardware token gate locking on-die SRAM in a 5th Protected (3'b100) state against weight snooping or overwrite.
omega_mollifier_core.sv: 66-cycle Q1.30 INT32 co-processor running in 264.0 ns (+/-0.4 ns at 250 MHz, 11,420 LUTs) with 0.00 B off-chip DRAM tape.
Public preprint & UVM 1.2 / LoadGen test logs on CERN Zenodo:
Put a bare-metal AMD Alveo FPGA node online wired to open-weights tool-calling agents so external AI safety researchers can try to bypass the 50.0 ms hardware cutoff over SSH.
Pre-silicon and HIL baseline already done: 134/134 formal assertions in Cadence JasperGold, 500 bare-metal FPGA runs, 50,000 LoadGen vectors matching FP64. Bench pass target: 100/100 injected out-of-bounds tool calls or kernel-hang faults physically clamped on an external oscilloscope within 50.0 ms with zero host OS help.
Open deliverables: MIT/Apache-2.0 release of the external AXI4 testbench wrappers, adversarial fault-injection scripts, and raw scope/logic-analyzer traces on GitHub (core proprietary .dcp checkpoint stays IEEE 1735 encrypted so safety teams can test the hardware black-box without exposing patent-pending RTL).
$1,000 minimum tranche: $450 SMA coax test cables & DC breakout headers, $350 PCIe 4.0/5.0 risers & high-speed loopbacks, $200 cloud synthesis compute.
$15,000 target tranche: adds $6,000 for hosted bare-metal AMD Alveo cloud instances for external red-team SSH access + $8,000 for open AXI4/SVA fault-injection wrappers and public documentation.
Up to $45,000 stretch tranche: funds dedicated FPGA evaluation kits shipped to independent AI safety red-teamers and red-team bug bounties against the 50.0 ms clamp.
Solo founder: Jonathan Ament, Charlotte, NC (https://www.linkedin.com/in/jonathanament | https://mavericksalphasignal.com). Recorded USPTO Provisionals 64/127,500 & 64/128,324 (Reel 075667 / Frames 0383 & 0567); active SAM.gov UEI RA2FZ1JUGCK8; DLA CAGE 26U94.
Main risks: (1) Python agent framework latency masking sub-microsecond bus behavior (mitigated via raw UDP/AXI4-Stream kernel-bypass framing); (2) safety researchers lacking local FPGA cards (solved by hosting remote bare-metal Alveo access).
Prior funding in last 12 months: $0 outside funding. Self-funded 100% of lab gear, FPGA runs, and patent filings out of personal savings.
There are no bids on this project.
Comments
No comments yet. Sign in to create one!