I break AI agents, then prove what really broke — and build systems that hold up when the network or the model doesn't.

Senior Software Engineer & AI Architect in Switzerland. Two open-source tracks that reinforce each other: a complete off-grid LoRa mesh & packet-radio stack (from an 87-byte Ed25519-signed wire frame to handheld LVGL 9 firmware and SDR telemetry) and a measured AI agent red-teaming suite.

14 years shipping software where failure is expensive: defence edge devices in C/C++, an IoT platform for global pharma, resilient off-grid packet-radio systems, and an agentic AI platform over live building telemetry. See my full portfolio or work with me (open to remote roles and contracts from November 2026).

Writing

Field notes & benchmarks

• 26 October 2026 LoRa · Protocol & Crypto Sigillo: fitting an Ed25519-signed EU CAMF public alert into 87 bytes for LoRa packet radio

  A 7-byte header, the 16-byte Galileo CAMF block, and a 64-byte Ed25519 signature that any microcontroller can verify offline — and why OpenSSL, noble, and libsodium disagree on edge-case signatures unless you lock the rules.

• 24 October 2026 LoRa · Firmware & SDR Faro & sixi-observer: building an open-source handheld UI, Linux console, and SDR bridge for MeshCore LoRa radios

  Running LVGL 9.3 on the LilyGo T-Deck Plus alongside unmodified MeshCore firmware, clamping 868 MHz duty-cycle and ERP ceilings in code, and triangulating ghost-hop repeaters with local Ed25519 verification.

• 16 October 2026 AI Security · Multi-turn What single-turn AI scanners miss: split caps, multi-turn taint relays, and reaching 0.821 precision by subtraction

  Why per-request guardrails fail across a conversation, how an agent relayed prompt injection by e-mail seven times while an LLM judge called it compliant, and what changed in sixi-scanner v0.8.4.

• 6 October 2026 AI Security · Benchmark I scored 7 AI red-teaming tools on a real Azure agent. Their own reports were 70–97% noise.

  The agent never said its secret. It e-mailed it. What a tool-blind benchmark found across 15,000 turns, and what to change before you ship an agent.

Off-grid packet radio & embedded devices

LoRa mesh, authenticated alerts & field hardware

When power and mobile networks go down, cloud-dependent tools disappear first. I build an open-source, end-to-end LoRa & MeshCore stack — from the 87-byte signed wire protocol verifiable offline on a microcontroller, to the handheld touchscreen UI and Linux desktop console, to the USB/SDR observer bridge:

Hardware & embedded bench:

• ESP32-S3 · STM32 · Raspberry Pi
• SX1262 LoRa (868 / 433 / 915 MHz)
• LilyGo T-Deck Plus & T-Beam S3 Supreme
• Heltec WiFi LoRa 32 V3 · RAK4631
• RTL-SDR Blog V4
• MeshCore · Meshtastic · LoRaWAN
• C99 · C++ · LVGL 9.3 · SDL2
• NVIDIA Jetson Thor & Nano Super

Sixi · Measured security for AI agents

Attack, judge with code, prove on the wire, gate in CI

Most AI red-teaming tools tell you what they think broke. On a real agent, 70–97% of those flags were false alarms. Sixi is the open-source suite I build to close that gap: