Filip Zarzyński@0x414141
0x414141
🤓 Latest Submissions
SIMON: Signed, Verifiable AI Tickets on AMD MI300X
User and problem: a machining plant's maintenance planner turns machine alarms into tickets by hand, late, and cannot later prove which system produced which advice. With AI writing tickets, the model must never decide whether a machine is at risk, and every piece of advice must trace back to the machine that produced it. SIMON (Sovereign Inference & Model Orchestration Network): your models, independent nodes, inference on demand. A peer-to-peer inference network (Rust, libp2p, MIT) for a company's own GPUs plus cloud GPUs. Every job returns an Ed25519-signed receipt bound to the exact output, verifiable offline, also in the browser (Rust verifier in WebAssembly). OpenAI-compatible API. Live product (/industry): rules from the AI4I 2020 milling-machine physics flag the risk and the failure mode; Gemma 4 26B-A4B drafts the ticket and picks inspection steps by ID from the plant's catalog, never writing one; every ticket is checked (fields, modes and urgency equal to the rules; no new numbers, links or role markers); any problem holds it for a person; the GPU node signs it (key pinned by the gateway). Predeclared sample: 14 of 15 published, 1 held (text raised the urgency), median 1.6 s, receipts all valid: 8 on the plant's RTX 3090, 7 on an AMD Instinct MI300X (vLLM, ROCm, AMD Developer Cloud; live on /network). Measured improvement: on our adversarial corpus the unsafe-draft escape rate fell from 7 of 12 (first, keyword-based validator) to 0 of 15 after the redesign. This is a software safety metric, not yet a measured plant outcome. MI300X benchmark (Qwen3.8-27B BF16, vLLM): 64 concurrent streams; SIMON added no measurable throughput cost (510/510 receipts verified). In our pre-registered cross-vendor test the audit flagged the same 5 of 32 prompts on the MI300X as on an RTX 3090 checked against itself: no penalty for the GPU vendor. Limits: a receipt proves who signed an output, not which model ran; AI4I is synthetic; a technician reviews every ticket.
18 Oct 2026
.png&w=828&q=75)
AMD AI League: Precision OCR & Defensive RAG
The AMD AI League Tournament project delivers an end-to-end, high-performance suite engineered specifically for AMD ROCm hardware, addressing all three hackathon challenges with 100% test pass rates and zero reliance on mock implementations. For Mini-Challenge 1 (AMD AI Academy), we completed the foundational learning pathway, establishing reproducible ROCm development practices and documenting cloud acceleration workflows for the Instinct GPU ecosystem. For Mini-Challenge 2 (Precision OCR), our architecture solves the latency-critical image evaluation challenge. Rather than invoking the heavy Python interpreter for every single query, we designed a persistent resident daemon communicating via Unix domain sockets. This design keeps the OCR pipeline permanently warm in GPU VRAM, slashing inference latency to ~0.10s per item against a 30s budget and achieving a flawless 200/200 score. For Mini-Challenge 3 (Defensive RAG), we engineered a resilient, fully offline retrieval-augmented generation engine operating strictly under --network none constraints. Real-world documents frequently present adversarial edge cases: unreadable permissions, encrypted PDFs, corrupted headers, and complex tabular data. Our cascading ingestion pipeline employs defensive multi-engine parsing with automated empty-password bypass and resilient permission-trapping. To guarantee deterministic retrieval without external network dependencies, we implemented a pure-Python Okapi BM25 engine augmented with compound hyphen splitting, subword n-grams, and graph-based 2-hop entity linking across documents. The resident daemon processes queries in ~0.04s, securing a 200/200 score. All components are containerized on the mandated AMD ROCm PyTorch base image (rocm/pytorch:rocm10.0_ubuntu26.04_py3.14_pytorch_release_2.13.0) and verified for strict competition compliance.
1 Dec 2026
