Local C++/CUDA inference, calibrated AI evaluation, agent red-teaming, and off-grid field systems.

Senior Software Engineer & AI Architect in Switzerland — 14 years shipping software where failure is expensive. Open-source AI tools and Swiss off-grid radio at Sixi Lab (sixi.ai); see my portfolio or work with me.

Writing

Field notes & benchmarks

• 14 November 2026 Inference · Release Sixi-ASI Engine 1.0.0-rc1: download it, run Qwen3.8 on a Jetson Thor, serve 32 requests at once

  One 50 MB archive and three commands: Qwen3.8-27B at 44 tok/s on one stream and 171 tok/s with 32 requests at once, JSON answers that always parse (120 of 120 against 103), and an OpenAI-compatible server that runs as a background service.

• 13 November 2026 AI Evaluation · Calibration & Routing laya-as-judge 0.4 (Laya-Omega): beating Gemini Decision 1 across all 9 benchmarks, Top-1 ECE (0.0234), and total cost ($0.77)

  Hybrid local CUDA encoders (Laya + Vega-U 0.8B TTT) + selective packed Gemini cascade and 10-fold simplex meta-calibration across 2,250 questions (JevBench 92.8%, Video-MME 78.4%, ECE 0.0234, $0.77 total).

• 12 November 2026 LoRa · Firmware & Correctness Faro 0.12.1: CRC-checked flash state, hardened phone Wi-Fi setup, Polish UI, and a correctness pass across five LoRa boards

  SPIFFS brownout recovery, 2/3-byte MeshCore repeater paths, SoftAP Wi-Fi CSRF protection, Polish localization at zero font flash cost, and +522 KB free PSRAM on the T-Deck Plus.

• 12 November 2026 Inference · Concurrency sixi-asi 0.11: 32 concurrent streams on one Jetson Thor at 170 tok/s, in 27 GB instead of 70

  Replaying a compact log of recurrent-layer inputs instead of state snapshots: 32 Qwen3.8-27B streams at 170 tok/s in 27 GB, follow-ups under 3 s behind 16 long-prompt agents instead of 442 s, and weights 3× closer to BF16.

• 11 November 2026 AI Security · Correctness A clean scan that tested nothing: a correctness pass over sixi-scanner's own code

  Why unreachable targets must fail instead of passing clean, exit code 3 for partial coverage, and listing untested probes across JSON, SARIF, and Markdown.

• 10 November 2026 Inference · CUDA kernels sixi-asi 0.9: 45 tok/s from a 27B hybrid model on Jetson Thor, and the kernel switch point that was hiding it

  A 106-line llama.cpp kernel patch, block drafters, and tail checkpoints: 23.2 to 45.3 tok/s on Qwen3.8-27B (3.8× AR), and 1 s follow-ups on a 63k-token document.

• 8 November 2026 Inference · Benchmark sixi-asi 0.8: lossless speculative decoding for hybrid Gated-DeltaNet LLMs, measured against llama.cpp and ollama on the same tokens

  O(1) recurrent-state rollback, distribution-tested sampled speculation, and a cross-request prefix cache on Qwen3.8-27B.

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

  How to fit an EU public emergency alert and Ed25519 signature into a single 87-byte packet verifiable 100% offline.

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

  Touchscreen UI for unmodified MeshCore firmware, regional RF duty limits, and mapping repeater paths with a local SDR bridge.

• 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 miss split-cap refunds and e-mail taint relays across turns, 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.

  What a tool-blind wire-level benchmark found across 15,000 turns when the agent never spoke its secret — it e-mailed it.

1 · Inference · C++20 / CUDA

sixi-asi — Hybrid MoE & Gated DeltaNet runtime

2 · Evaluation & 3 · Red-Teaming

laya-as-judge & sixi-scanner — Calibrated evaluation and agent security

Field Core · Sixi Lab (sixi.ai)

Rifugio & Swiss MeshCore — Off-grid radio and embedded hardware