Hardware & embedded
engineer.

Modern cars are heavily sound-insulated and their stereos are loud, so drivers often notice an approaching ambulance, police car or fire truck late — or not at all. Deaf and hard-of-hearing drivers cannot hear the siren in the first place. SirenSense solves this at the source: it detects the siren acoustically, classifies it on-device, computes its direction relative to the car, and warns the driver with a direction arrow and a spoken cue ("siren, left") in time to react.

How it works

The system is three linked units. A roof node carries a four-microphone array in a 190 mm cross; a convolutional neural network trained with Edge Impulse decides "siren / not siren" on-device, and a time-difference-of-arrival (TDOA / GCC-PHAT) calculation turns the tiny delays between microphones into a bearing. The bearing is sent over ESP-NOW to a cabin node, which draws the arrow, speaks the alert, logs the event with GPS, and listens on a LoRa radio for transponders in emergency vehicles up to ~2 km away. A Raspberry Pi HUD renders the driver display and an offline map.

Engineering worth noting

• Two custom PCBs designed in KiCad (roof node, cabin node) plus a transponder board, all fabricated at JLCPCB under my Tekkera Electronix label.
• Dual-I2S audio capture at 44.1 kHz across two synchronised buses feeding the four mics.
• A real debugging story: the radio link went silent in the car while both nodes reported "send OK". Both ESP32s had left Wi-Fi power-save on, which silently drops ESP-NOW broadcast frames — one line, esp_wifi_set_ps(WIFI_PS_NONE), fixed it.
• An encrypted LoRa transponder network (AES-128) so emergency fleets can pre-announce themselves 30–60 seconds before they are in sight.

The version you can see on the site is the working prototype that won Hack4Change and was submitted to the James Dyson Award. The three phases below trace the build from a breadboard to that unit.

SUBJECT 105 is a handheld RF/hardware audit tool in the spirit of the Flipper Zero, redesigned to add the things a stock Flipper lacks — on-board Wi-Fi, a 2.4 GHz research radio, GPS and a sunlight-readable display — while staying legal for EU operation. It is a portfolio project that puts almost every skill a hardware employer screens for into one object you can hold.

Architecture

It is dual-processor: an STM32WB55 (Cortex-M4 with integrated BLE) runs the real-time RF work and the UI, and an ESP32-S3 co-processor adds Wi-Fi. They talk over a custom framed UART/SPI protocol (CRC-16, request/response plus async events) designed to be byte-identical on both sides. The radio front-ends include a CC1101 (sub-GHz), an ST25R3916 (NFC), a 125 kHz stage, IR, an nRF24L01+ and a u-blox GPS, behind a FreeRTOS app/plugin firmware where each chip is a driver behind a clean interface and the only vendor-specific code is one HAL shim.

Legal & ethical by design

This is the part I care about most. The device is a security-research and RF-education instrument. Firmware enforces EU band, power and duty-cycle limits, and it deliberately does not implement jamming, Wi-Fi deauth/beacon-flood, credential phishing or arbitrary RF injection — those are excluded on both processors, not merely undocumented. Where attack awareness matters, it implements the defensive side: deauth, evil-twin and beacon-spam detectors.

The project is in active development.

Copilot Pro is not a rewrite of SirenSense — it is a new Pi-side application that talks to the same two ESP32 nodes over the same UART JSON protocol, reuses SirenSense's safety modules unchanged, and adds an infotainment layer on top. It is aimed at my own 2015 Citroën C4 Cactus, whose stock screen is slow and lacks CarPlay / Android Auto.

What it adds

• Eight touch pages in pygame at native resolution — home, the original safety screens, a full-res offline map, animated OBD-II gauges, media, cameras, a system view and settings.
• OBD-II live engine data (RPM, speed, coolant, intake, battery, load) over Bluetooth ELM327.
• Bluetooth audio out to the car's stock head unit — the Pi pairs like a phone.
• Rear camera (USB UVC), WLED ambient lighting that follows the alert priority, and a full-screen crash / fatigue / threat takeover system shared with the base product.

A key design rule: the Pi never touches the car's CAN bus. It connects only through ignition 12 V, Bluetooth audio, and the standard OBD-II port. The software is complete and screenshot-verified across all eight pages; it is not yet installed in the car.

The prototype proves the idea; V2 is about making it a product. The plan moves the three safety "brains" from ESP32 to automotive-grade STM32 — for temperature grades, long-term availability, determinism and an easier path through EMC and E-mark certification. The roof pod gains automotive MEMS microphones and BLE Coded PHY in place of ESP-NOW; the cabin hub adds CAN-FD; and the whole thing runs on Zephyr RTOS, keeping the LoRa AES protocol, the microphone geometry and the alert-priority model.

It is positioned deliberately as driver alerting, not intervention, to stay clear of the heaviest automotive-safety standards, with a certification path through RED, UNECE R10 E-mark and UN 38.3. The real long-term moat is a per-country siren dataset. This is a vision and an architecture — honestly, not yet built — but it is the direction the work is heading.