Hi, I’m Ben.

ECEN 3730 · PCB Design · Fall 2026

555 timer board: design, bring-up & characterization

A business-card-sized board with an LMC555 driving four LEDs through 10 kΩ to 50 Ω resistors. It worked on first power-up with no rework. I brought it up in stages using jumper isolation switches and characterized it with a scope.

• 502 Hzmeasured vs. 481 predicted
• 66%duty cycle (66.7% designed)
• ~32 Ωextracted 555 Thevenin R

• Altium
• Bring-up
• Oscilloscope

ECEN 3730 · Labs 11 & 12 · Fall 2026

Trace resistance & current capacity

I measured eight copper traces four different ways and showed that 2-wire readings were 2 to 40× too high because of lead and contact resistance. Then I pushed current through a 5 mil trace until it smoked and compared the result with the IPC-2221 limit.

• 0.41 mΩ/sqmeasured sheet R
• ~1.6 milinferred copper thickness
• 5 Ato smoke (IPC: ~1 A)

• 4-wire Kelvin
• IPC-2221
• Data analysis

Project · Altium Designer

STM32 board with Bluetooth & USB-C

A 4-layer STM32 board that adds USB-C data and BLE telemetry for wireless diagnostics during testing. I worked with three other engineers on the stack-up and high-speed routing constraints to keep EMI down, and wrote oscilloscope test procedures that sped up bring-up.

• 480 MbpsUSB-C data
• 4-layercontrolled routing
• +20%debug efficiency

• Altium
• High-speed layout
• USB
• BLE

ECEN 4610 · Senior Capstone · Fall 2026 · in progress

SPICE: photon-statistics timing firmware

This is the firmware for a Hanbury Brown–Twiss setup. It timestamps every pulse from two single-photon detectors on one shared clock and builds the g²(τ) histogram that tells you whether light is random, bunched or made of single photons. I run the STM32H723 at 550 MHz with dual-channel timer input capture streamed into circular DMA buffers.

• 3.6 nstimestamp resolution
• 101 × 10 nshistogram bins
• 64-bitwrap-extended time base

• STM32H7 / Cortex-M7
• DMA
• Input capture
• Cache coherency

ECEN 5003 · Embedded AI (graduate) · Fall 2026

Keyword spotting on an STM32F407

A wake-word pipeline that runs entirely on the microcontroller. It goes from the PDM microphone through I²S/DMA to an STFT spectrogram, then a small CNN running in ExecuTorch, with predictions streamed over UART. I trained a compact model that beat the 11-million-parameter reference while fitting comfortably in 1 MB of flash.

• 95.3%validation accuracy (ref. 93.3%)
• ~1,100×fewer parameters
• 1.2 send-to-end on board

• PyTorch
• ExecuTorch
• DSP
• C++
• Renode

ECEN 5003 · Embedded AI (graduate) · Fall 2026

Neural-network inference on bare metal

I trained a 1-16-16-1 tanh network to approximate sin(x), exported it to ExecuTorch, and ran it on a Cortex-M4 with no OS to drive LED brightness through PWM. I trimmed the kernels down to three operators, put the memory pools in CCM RAM, and found a UART baud bug on macOS by reading fault registers over SWD.

• 0.00045mean error vs. sinf()
• 4.1 KBmodel
• 7.5%flash used

• ExecuTorch
• CMake
• PWM
• SWD debugging

ECEN 3593 · Computer Organization · Spring 2026

Pipelined RISC-V CPU in SystemVerilog

I built a 5-stage RV32I processor over seven phases: an instruction decoder, a pipelined datapath with a register file, ALU and immediate instructions, data forwarding to resolve hazards, branches, and finally loads and stores. I verified each phase with hand-written RISC-V assembly tests in Icarus Verilog and GTKWave.

• SystemVerilog
• RISC-V assembly
• Icarus Verilog
• GTKWave
• Hazards & forwarding

Personal project · iOS · 2026

Slice: scan a receipt, split the bill

A SwiftUI app that scans a receipt with the document camera and parses the line items. You assign each item to a friend from your contacts, and the app splits the total. Receipt parsing runs through a Firebase Cloud Functions proxy that reconciles an on-device heuristic parse with an LLM parse. Payments go through Stripe. Built for TestFlight.

• Swift / SwiftUI
• VisionKit
• Firebase Functions
• Stripe