https://digicrafter.ai/blog.html Long-form field notes from the DigiCrafter studio. Production lessons, dead ends, and the occasional opinion. en-us Wed, 22 Apr 2026 10:00:00 +0000 https://digicrafter.ai/blog-7gpu-vulkan-fleet.html https://digicrafter.ai/blog-7gpu-vulkan-fleet.html Wed, 22 Apr 2026 10:00:00 +0000 Infrastructure We built a local AI inference rig out of used mining hardware and consumer AMD cards. Total spend: about $2,000. Here's why we chose AMD, why Vulkan over ROCm, what broke, and the one lesson we kept rediscovering until we finally wrote it down. https://digicrafter.ai/blog-section-aware-chunking.html https://digicrafter.ai/blog-section-aware-chunking.html Sat, 18 Apr 2026 10:00:00 +0000 Retrieval For three weeks our recall on a 50-document GST validation set was stuck at 0.51. Naive recursive chunking does not work on tax law. The fix wasn't a fancier embedder — it was admitting that the chunker had to understand legal structure before it tokenized anything. https://digicrafter.ai/blog-llama-cpp-vs-ollama.html https://digicrafter.ai/blog-llama-cpp-vs-ollama.html Sat, 28 Mar 2026 10:00:00 +0000 Pragmatism Ollama is great for desktop chat. It is not great for ingesting 14,000 PDFs at six chunks per second per GPU. Here's the short version, including the table of "what we tried vs. what failed" so you don't have to redo it. https://digicrafter.ai/blog-local-first-ai.html https://digicrafter.ai/blog-local-first-ai.html Sat, 14 Mar 2026 10:00:00 +0000 Manifesto Most AI consultancies build you a wrapper around an OpenAI key. Six months in, you're locked in, your data lives in their logs, and your costs scale with every user you add. Here's when local-first is the right call, when it isn't, and the cost math that surprised us.