Work with me

I help teams run open-weight AI models faster on local hardware, replace noisy LLM judges with calibrated evaluation, test AI agents before they ship, and engineer off-grid edge and radio systems. Remote, from Switzerland.

1. Local LLM & Hybrid MoE inference on NVIDIA Jetson and DGX Spark

• Native C++20/CUDA inference for hybrid MoE and linear-attention models (sixi-asi) with strict memory caps, lossless speculative decoding, input-log replay, and zero Python on the runtime path
• Deployment on NVIDIA Jetson (Thor, Orin, Nano Super) and DGX Spark unified memory, plus OpenAI-compatible HTTP/SSE serving and schema-checked tool calls

Proof: on Jetson AGX Thor with Qwen3.8-27B, 45.3 tok/s single-stream (3.8× AR), 32 concurrent streams at 170 tok/s in 27 GB instead of 70 GB, and 1.0 s follow-ups at 63k context.

2. Calibrated AI evaluation & secure agent platform architecture

• Hybrid local + cloud evaluation (laya-as-judge): 7–14 ms local GPU encoders at $0.00 + selective cloud escalation and simplex probability calibration
• Enterprise agent architecture on Microsoft Foundry or Google ADK: Entra agent identities, scoped MCP/A2A tools, governed AI gateways, and EU AI Act / DORA / GDPR evidence

Proof: laya-as-judge v0.4.0 wins 11/11 benchmarks over Gemini Decision 1 across 2,250 questions (Top-1 ECE 0.0234, $0.77 total); 14 years across defence, pharma, and building-automation platforms.

3. AI agent security assessment & CI red-teaming

• Wire-level red-teaming over REST, MCP, and A2A — including multi-turn state abuse and indirect injection through tools and knowledge bases
• sixi-scanner v0.9.1 wired into CI with native SARIF 2.1.0 output and custom deterministic business-rule oracles

Proof: 7-tool benchmark and multi-turn taint-relay write-ups — 0.821 stateless precision (1st of 7) and 1.000 multi-turn precision.

4. Off-grid LoRa mesh & embedded field systems

• Autonomous off-grid MeshCore infrastructure (Sixi Lab / Rifugio) and compact offline-verifiable signed radio protocols (sigillo v1.1-rc: 87-byte Ed25519-signed EU CAMF alerts)
• Embedded C/C++ and LVGL 9 firmware across five ESP32-S3 LoRa boards (Faro v0.12.1), SDR telemetry (sixi-observer, map.sixi.ai), and Jetson edge diagnostics (GlassBoxEdge)

How it starts

A 30-minute call on what the system does, what hardware or data it touches, and what would hurt — followed by a fixed-scope proposal. Email contact@sixi.ai.