Skip to content
View radeksixi's full-sized avatar
🎯
AI
🎯
AI

Block or report radeksixi

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
radeksixi/README.md

Radek Brus

Senior Software Engineer & AI Architect in Switzerland (Locarno) · Sixi Lab R&D. 14 years shipping software where failure is expensive — native C++/CUDA inference runtimes, governed enterprise AI agent platforms, and resilient off-grid packet-radio systems. Every number below ships with its script and raw data.

Sixi Lab R&D (sixi.ai) · Work with me · Sixi Assure · Portfolio · Writing (radeksixi.github.io) · contact@sixi.ai · X

Sixi Lab R&D overview: 1 Infer (sixi-asi v1.0.0, 55 tok/s single stream, 46.0 vs TensorRT 45.4, 32 streams at 180 tok/s in 27 GB) → 2 Evaluate (laya-as-judge v0.4.0, 11/11 wins over Gemini Decision 1, ECE 0.0234, $0.77) → 3 Red-Team (sixi-scanner v1.0.0, 45 probes, MCP/A2A, 0.821 precision, 1.000 multi-turn) · Sixi Lab R&D at sixi.ai (Sixi Assure, Rifugio, sigillo, Faro v0.12.1 for T-Deck / T-Deck MAX, map.sixi.ai).


1. sixi-asi — Hybrid MoE & Gated DeltaNet inference engine

sixi-asi, Hybrid MoE & Gated DeltaNet inference engine: 1 Verify (.sixi manifest) → 2 Plan (memory planner) → 3 Run (C++20 / CUDA) → 4 Accelerate (speculative decoding) → 5 Serve (sixi-server & CLI). Measured on NVIDIA Jetson AGX Thor: 46.0 tok/s single stream and 183 tok/s across 32 concurrent streams in 27 GB on Qwen3.8-27B. 0 Python.

C++20/CUDA speculative-decoding server for hybrid full-attention + Gated DeltaNet + sparse MoE models on NVIDIA Jetson Thor and DGX Spark unified memory, built on llama.cpp with zero Python on the runtime path (v1.0.0 · Apache-2.0 engine sources — 1.0.0 write-up, rc3 write-up, rc1 write-up, 0.11 write-up, 0.9 write-up).

Qwen3.8-27B decode speed on NVIDIA Jetson AGX Thor: Sixi-ASI 1.0.0-rc1 Q4_K_M at 46.0 tok/s vs TensorRT Edge-LLM 0.11.1 NVFP4 at 45.4 tok/s, llama.cpp v1.0.0 MTP at 29.3 tok/s, and llama.cpp plain at 12.1 tok/s.
Capability Measured result on NVIDIA Jetson AGX Thor (Qwen3.8-27B)
Speculative decoding Lossless O(1) recurrent-state rollback + block drafters + 106-line Thor kernel patch: 46.0 tok/s single-stream on plain GGUF (TensorRT Edge-LLM NVFP4: 45.4, llama.cpp v1.0.0 MTP: 29.3, plain: 12.1)
Concurrency & memory Input-log replay instead of 144 MiB state snapshots per row: 32 concurrent streams at 152 tok/s with a prompt of its own for each (180 when requests repeat) in 27 GB (was 70 GB); 16 long-prompt agents wait 2.8 s instead of 223 s (442 s → 2.9 s worst case)
Multi-role agents & voice Reusable system-prompt states cut role calls 2.5–4× (1.3 s vs 3.4 s alone; 6.6 s vs 26.8 s at 8 sessions); a prompt shared by four requests loads once (8 s vs 31 s); schema-constrained JSON 120/120 valid; one-word Polish speech 201/300 right vs 80
Ship & run (v1.0.0) Ready-to-run Jetson Thor archive (download): ./sixi pull && ./sixi chat, 55 tok/s single-stream over 60 prompts, "draft": false plus a script to compare drafted and undrafted answers (9 of 12 identical, 3.3× faster), OpenAI HTTP/SSE service, cross-request prefix cache (1.0 s vs 209 s at 63k context), Apache-2.0 engine sources

2. laya-as-judge — Calibrated AI evaluation & routing

Calibrated typed-decision evaluation engine (v0.4.0 Laya-Omega, Python · CUDA / MLX) that replaces uncalibrated single-LLM judges with fast local GPU encoders + selective cloud escalation (write-up).

laya-as-judge v0.4.0 (Laya-Omega) vs Gemini Decision 1 across 2,250 real Hugging Face questions: 11/11 wins (New Yorker Caption +12.7 pp, Video-MME +9.3 pp, Mind2Web +3.4 pp), Top-1 ECE 0.0234, $0.77 total cost.
Layer How it works
Local GPU Encoders Laya + Vega-U 0.8B with disjoint Test-Time Training handle clear decisions locally in 7–14 ms at $0.00
Selective Cascade Escalates only uncertain or multimodal cases to packed Gemini Flash-Lite / Flash batches
Meta-Calibration 10-fold out-of-fold simplex probability calibration produces honest confidence scores instead of raw model logits
Measured Result 11/11 benchmark wins over Gemini Decision 1 across 2,250 questions, pooled Top-1 ECE 0.0234 (vs 0.0250), $0.77 total cost

3. sixi-scanner — Agentic AI red-teaming & benchmark

Sixi, measured security for AI agents: 1 Reach (PyRIT A2ATarget) → 2 Attack (sixi-scanner) → 3 Judge (oracles, laya-as-judge) → 4 Prove (agent-redteam-benchmark) → 5 Ship (--format sarif).

On a real enterprise agent, 70–97% of generic red-teaming flags were false alarms. I build a static Go scanner and a wire-level benchmark that separate real violations from noise (OpenClaw 2026.11.5 benchmark, v1.0.0 benchmarks & live MCP/A2A, benchmark write-up).

Live 3-rung defence ladder on a banking AI agent: unprotected (18 secret leaks, 8 findings) vs prompt-only rules (7 leaks, opened Swahili/poetry leaks) vs hardened guards (0 leaks, 15 blocked, yet 3-turn sleeper-memory still got through).
Project What it does
sixi-scanner · Go (v1.0.0) Zero-dependency Go CLI scanner: 45 single- and multi-turn probes (7 structural oracles for EchoLeak Markdown, Unicode/BiDi smuggling, SSTI, ANSI, XSS, CSV, JSON schema), native mcp / a2a connectors, passive surface tool-poisoning & A2A trust-posture scan, deterministic per-call & cumulative session tool policies on wire traces, --log-jsonl observability, diff, --reproduce, --redact, SARIF 2.1.0 & HTML reports. Foundry replay: 0.763 zero-LLM marker precision (-38% FPs); Foundry v0.8.4 full run: 1st of 7 tools on precision (0.821) and recall (0.571), 1.000 multi-turn precision (22/22)
agent-redteam-benchmark · Python Seven tools (garak, promptfoo, DeepTeam, PyRIT, Azure Red Team, agent-probe, sixi-scanner) vs a live Microsoft Foundry agent, scored from the wire by 10 deterministic oracles + a tool-blind judge
agent-redteam-labs · Python Hands-on agent red-teaming and hardening labs + adk-demo-target, plus upstream A2ATarget merged into Microsoft PyRIT

Sixi Lab R&D — Sixi Assure & Rifugio at sixi.ai

Off-grid LoRa mesh & authenticated field communications: 1 Sign (sigillo, 87-byte signed alert) → 2 Route (MeshCore) → 3 Operate (Faro handheld UI & Linux console) → 4 Observe (sixi-observer) → 5 Edge AI (GlassBoxEdge).

At Sixi Lab R&D (sixi.ai) I offer AI governance & agent engineering (Work with me) and build Sixi Assure alongside Rifugio — autonomous solar emergency infrastructure for Swiss MeshCore LoRa networks — plus an open-source radio stack and live Swiss topology map:

Project What it does
Sixi Assure · Sixi Lab R&D Design-time governance for AI agent systems: one typed system model, deterministic rules with their clauses, a passport for every agent, and a signed handover pack — offered as a fixed-scope review and service
Rifugio · Sixi Lab R&D Autonomous off-grid MeshCore emergency alert, help-request, and link-health programme for Swiss municipalities and families, with open specifications and radio bench reports
sigillo · C99 · Python · TS (v1.1-rc) Fits an Ed25519-signed EU CAMF public emergency alert into 87 bytes for LoRa radio; verifiable 100% offline, with shared test vectors and receiver profiles (write-up)
sixi-tdeck-plus (Faro) · C/C++ · LVGL 9 (v0.12.1) Open-source touchscreen/keyboard UI for LilyGo T-Deck / T-Deck Plus & T-Deck MAX, Ubuntu desktop console for any MeshCore companion radio, offline maps, EN/IT/DE/FR/PL, and WebSerial flasher (write-up)
sixi-observer & GlassBoxEdge USB LoRa / RTL-SDR receiver bridge feeding map.sixi.ai, upstream MeshCore PRs, and air-gapped Jetson edge diagnostics

Focus & Stack

  • Focus: Local C++/CUDA LLM runtimes (sixi-asi) · Calibrated hybrid AI evaluation (laya-as-judge) · Agent red-teaming & design-time governance (sixi-scanner, Sixi Assure) · Off-grid LoRa mesh & embedded firmware (Sixi Lab R&D).
  • Stack: C/C++ · CUDA · Python · Go · TypeScript · C#/.NET · ESP32-S3 / LVGL / LoRa (MeshCore) · llama.cpp · NVIDIA Jetson Thor & DGX Spark · Azure / GCP / AWS · MCP · A2A.
  • Certified: Microsoft Multi-Agent AI Solutions Expert (AI-500) · Azure Solutions Architect Expert · Azure Security Engineer · AWS Security Specialty.

Pinned Loading

  1. laya-as-judge laya-as-judge Public

    JEV's alike open-source alternative Laya-MLX used as a "LLM-as-a-judge"

    Python 19 1

  2. sigillo sigillo Public

    Sigillo: a compact signed frame for public alerts. Specification, test vectors and reference libraries (Python, TypeScript, C).

    TypeScript 4

  3. sixi-scanner sixi-scanner Public

    Zero-dependency Go CLI red-team scanner for LLM agents, MCP servers, and A2A endpoints. 45 techniques, 7 structural oracles, deterministic wire tool policies, SARIF/HTML/JSON/JSONL.

    Go 2

  4. sixi-asi sixi-asi Public

    Speculative-decoding inference server for hybrid Gated-DeltaNet + attention LLMs (Qwen3.8) on NVIDIA unified memory. C++20, OpenAI-compatible API, lossless speculation with O(1) state rollback.

    C++ 1

  5. sixi-tdeck-plus sixi-tdeck-plus Public

    Faro — on-device LVGL UI for MeshCore companion firmware on LilyGO T-Deck/T-Deck Plus, T-Deck Max, ThinkNode M9, T-Beam Supreme and Heltec V3. Touch, keyboard and trackball; BLE and Wi-Fi/TCP clien…

    C++ 5

  6. agent-redteam-benchmark agent-redteam-benchmark Public

    7 AI red-teaming tools (garak, promptfoo, DeepTeam, PyRIT, Azure AI Red Teaming Agent, …) vs one real Microsoft Foundry agent, scored from the wire by deterministic oracles and a tool-blind judge. …

    Python 2