Custom AI for e-commerce — and a publication about running large language models on hardware you own.
We build AI sales assistants that know your products, pricing, and inventory in real time. We also measure things and publish the numbers, including the ones that don't flatter the hardware.
🌐 ai.rs · 📧 github@ai.rs · 📍 Belgrade, Serbia
Everything below is measured on our own hardware, with the methodology in the article.
| DeepSeek-V4-Flash on Two GB10s | 304B params, 1M context, 88 tok/s on 83 W — and a CUDA kernel it got wrong |
| Self-Hosted Claude Code: What Each Memory Tier Buys | 32 GB, 128 GB, 256 GB — what each actually runs |
| RTX 5090 vs GB10 (HP ZGX) | 32 GB at 1.8 TB/s against 128 GB at 273 GB/s |
| Will This LLM Fit My GPU? | The VRAM formula, and a one-command check |
| 4-Bit Quantization Decoded | INT4 QAT, MXFP4, NVFP4 — what the formats actually change |
| Qwen 3.6 27B: a Local Coding Model You Can Actually Run | Single 4090, within 4 points of Opus 4.6 on SWE-bench |
| Qwen3-Coder 30B-A3B on RTX 5090 with Ollama | 231 tok/s, 64K context, full setup log |
| Gemma 4 LoRA Fine-Tuning on RTX 5090 | What works, what doesn't |
| Synthetic Data for Fine-Tuning | Generating your own training set |
| Mixture of Experts, Explained | Why 304B params can activate 13B per token |
| The KV Cache | The trick that makes LLMs fast, then slow |
| Prompt Processing vs Token Generation | Two speeds, and why one number misleads |
| What a 256K (or 1M) Context Window Costs | Memory, compute, and bandwidth per token |
| What is RAG and Why Your AI Needs It | Retrieval, grounded in a real use case |
| LLM Post-Training Explained | SFT, DPO, and GRPO |
| Gemma 4 vs Qwen 3.5 vs Llama 4 | Updated benchmarks |
| Kimi K2.6 Explained | A trillion-parameter open model, and its memory bill |
| Claude Opus 5: What's New | What changed, and whether to switch |
| BM25 vs Embeddings | Why keyword search still wins for product search |
| LLM Security: From 17% to 94% Attack Resistance | Prompt injection, measured before and after |
New to this? The site has a beginner track with no code and no jargon.
- gb10-deepseek-v4-flash — two-node DGX Spark operator tooling and measurement harness for DeepSeek-V4-Flash: launch and status scripts, 2 Hz GPU-rail power sampling, and a harness that executes generated code against hidden tests.
- LLM fine-tuning — custom-trained open models on your product catalog
- RAG pipelines — live retrieval from your database for accurate pricing and availability
- AI sales assistants — embedded chat that recommends products and answers questions
- Dedicated deployment — your own server, your own data, no shared infrastructure
Stack: fine-tuned open-weight LLMs (Qwen, Gemma, Llama, DeepSeek), custom RAG with vector search plus live database queries, Rust and Python infrastructure, on-premise or dedicated cloud.
Serving e-commerce businesses across Europe and beyond.