Hello, I'm a grad student at Johns Hopkins. I build digital hardware and then try to prove it wrong.
Most of what's here is SystemVerilog checked against a reference model, formal proofs and real hardware when I have it.
Right now I'm adding a controller to my systolic-array accelerator, and next I'm putting a UVM testbench on it. Below are the projects that are done and verified.
| Project | What it is | Evidence |
|---|---|---|
| tensor-mac | INT8 systolic-array matrix multiply | Matches NumPy on 500 random tiles; formal proof that the accumulator never overflows |
| verified-rtl-blocks | FIFO, skid buffer, round-robin arbiter | 7,500+ random checks; formal proofs of no overflow and no starvation |
| fpga-keyword-spotting | int8 neural net in VHDL on an Artix-7 | Bit exact with Python on 128/128 vectors, on the board |
| risc8-rtl-to-gdsii | 8-bit CPU from Verilog to 45 nm layout | Six-corner sign-off at 7 ns with +1.24 ns slack |
| Radiation-hardened SRAM | Reproduced a published 16T cell, then removed 3 transistors | Read SNM +33%, hold SNM -64% |
Tools: SystemVerilog, VHDL, Verilator, Icarus, cocotb, SymbiYosys, Yosys, Vivado, Cadence Genus / Innovus / Xcelium, Python, C++.
