Essays, field notes and technical deep dives from the engine room. On sovereign AI, on-premise deployments and the tools we build with.
Mandatory time tracking is both burden and opportunity. Implementing it with an AI agent now eliminates both: the manual typing and the gaps in cost transparency.
Coding assistants no longer have to live in the cloud. Three hardware approaches for SMB teams — DGX Spark, ASUS GX10, multi-RTX 6000 Pro — and the models that actually run productively on each.
After three years of cloud-first, the counter-movement is here: companies are pulling AI workloads back into their own data centers. Privacy, cost, latency — and a clear trend for the German Mittelstand.
A productive LLM system needs more than a model. Guardrails, evals and red teaming decide whether a demo turns into a reliable long-runner.
RAG becomes productive when search is precise. How hybrid search, reranking and GraphRAG bring answer quality to the level enterprises actually need.
From LLM prototype to productive system: LLMOps defines the disciplines separating demo from sustained operation. What belongs in the stack in 2026.
Not every problem needs 70 billion parameters. When small language models are the better choice — more private, faster, cheaper, local.
Behind every modern LLM stands an architecture called the Transformer. How attention works, why it changed everything, and which building blocks really matter.
Multimodal AI crossed the language threshold in 2026: images, documents, audio are daily reality. How VLMs work and where they actually create value.
Tokens cost money, context windows define possibilities. How tokenization, long context, and retrieval memory interact in productive LLM systems.
From pretrained model to helpful assistant: instruction tuning, RLHF and DPO. Which method fits when — and how to align models to your own preferences.
Open-source LLMs grew up in 2026. When they're the better choice than closed-source APIs — and which aspects are commonly overlooked.
PEFT isn't synonymous with LoRA. Which parameter-efficient adaptation methods matter in 2026 and what trade-offs they make on effort, quality and modularity.
Synthetic training data isn't a stopgap anymore but a productive material. When it helps, when it hurts — and how to use it cleanly.
LLMs become productive once they can drive software. How tool calling, function calling and MCP make that possible — and which security questions come with it.
LLM inference isn't magic, it's math under memory pressure. What drives the speed and cost of AI answers — and how to optimize both.
Embeddings translate meaning into vectors. Vector databases retrieve them. Why these two building blocks are the invisible foundation of almost every modern AI application.
When the big model is too expensive for production: distillation turns a 70B teacher into a 7B student — often at 90%+ quality and drastically lower cost.
Mixture-of-Experts decouples capacity from compute: many experts, few active per token. How Mixtral, DeepSeek and others benefit.
QLoRA makes fine-tuning a 70B model on a single GPU possible. How 4-bit quantization pulls that off — and where the limits are.
Fine-tuning is worth it less often than marketing decks claim. When it really makes sense — and which preconditions must be in place to keep it from failing expensively.
LoRA makes LLM fine-tuning cheap: only small adapters are trained, the rest of the model stays untouched. How it works and when it's the right choice.
Which method actually brings an LLM to your use case? A clear side-by-side of RAG, fine-tuning, LoRA, prompt engineering and distillation — with a decision matrix.
Classic software testing isn't enough for modern AI. How to seriously evaluate frontier models and production applications in 2026 — and keep risk under control.
Three years after the first hype cycle, clear patterns have emerged. Which building blocks survive in sovereign AI deployments — and which ones were marketing.
From computer to lab and back: how AI predicts protein structures, designs drugs and discovers materials — and what that means for in-house R&D.
We understand how an LLM is built. We often don't understand why it does what it does. Mechanistic interpretability tries to close that gap.
An LLM knows words; a world model knows physics. Why the next generation of AI is learning to simulate the world — and what that changes in practice.
Reasoning models don't answer right away — they think first. What sits behind the o-series, DeepSeek-R1 and Gemini Thinking, and where it's actually useful.
RAG looks fantastic in a POC. In production it burns trust — quietly, for months. Seven patterns we keep seeing.
Rare, long-form, no tracking. We write when we have something real to say — not more often.
