Blog
Short notes from the field — solutions, problems, lessons learned. We write when we have something to say, not when a deadline says so.
May 25, 2026 · 10 min read
On RAG deployments clients always first tune the prompt and the model. The reality: 80% of the gap between a pipeline at 60% and 90% precision comes from chunking, embedding model and reranker. Concrete numbers.
May 11, 2026 · 10 min read
With H100 SXM5 + B200 + GB200 there's a moment when air cooling physically ends. When DTC, when immersion, why two-phase is dead, and what a cluster with PUE 1.08 really costs.
April 29, 2026 · 9 min read
Shop-floor AI copilot is the best-selling slide for manufacturing in 2026. Out of 10 deployments we've seen 4 that paid back and 6 that burned money. Concrete numbers on where the difference lies.
April 16, 2026 · 10 min read
Classic RAG works well as long as questions are straightforward. Agentic RAG lets an agent decide what to retrieve, when, and how many times — with a real cost premium attached.
March 30, 2026 · 10 min read
You don't need to pay for a frontier model on every request. LLM routing classifies query complexity and sends each request to the cheapest model that can handle it — the expensive one only gets what it truly needs.
March 18, 2026 · 10 min read
Industrial technicians spend hours searching thousands of pages of manuals and standards. LLMs with RAG change that — but only if you know where the naive approach fails and where it actually works.
March 2, 2026 · 9 min read
80% of "we want to deploy AI on our corporate database" projects begin with the question "do we fine-tune the model on our data?". The answer in 90% of cases is no. RAG is cheaper, faster to update, better to audit, and usually gives higher-quality answers.
February 17, 2026 · 10 min read
Frontier models score 88–94% on MMLU and the benchmark has lost all discriminative power. Learn which numbers actually say something about your use case — and why your own eval always matters more than any leaderboard.
February 5, 2026 · 10 min read
Generating training data with a stronger model sounds like a shortcut — and sometimes it genuinely is. But the same approach can lock a model inside a loop of its own errors, with no way out short of a full restart.
January 19, 2026 · 10 min read
Pure vector retrieval fails where exact wording matters — standard numbers, product codes, paragraph references. Hybrid search combines BM25, dense vectors and a reranker, pushing recall up by 10–15 percentage points.
January 6, 2026 · 10 min read
Dozens of LLM models compete at the same price point today. How do you pick the right one for a specific use case — without wasting time on unnecessary tests or overspending?
December 19, 2025 · 11 min read
Prompt injection has sat at #1 on the OWASP Top 10 for LLM applications for years — and no prompt or system filter can eliminate it entirely. We show how to layer your defences so the risk stays manageable.
December 8, 2025 · 10 min read
An AI agent without human oversight is powerful — and dangerous. When to require approval, when to trust it, and how to gradually expand autonomy without losing control.
November 26, 2025 · 10 min read
An agent without guardrails is not a product — it's a prototype waiting for an incident. We show how to layer your defences from input validation through action allowlists to a kill-switch.
November 7, 2025 · 10 min read
Buy an off-the-shelf AI tool or build a custom one? Most companies answer this question too fast — and pay for it with vendor lock-in, or years of wasted pilots.
October 27, 2025 · 11 min read
Continuous batching, KV-cache, and PagedAttention determine whether you extract 5 or 50 parallel requests from a single GPU — and whether your LLM serving is economically sustainable or not.
October 10, 2025 · 10 min read
Fine-tuning without evaluation is a gamble — you don't know what you gained or what you lost. Here is a systematic approach to measuring the real benefit and detecting catastrophic forgetting before production does it for you.
September 29, 2025 · 10 min read
LangGraph, CrewAI, AutoGen, or plain Python — the framework you pick determines what you get in production and what will cause you pain. Here is a practical decision framework.
September 16, 2025 · 10 min read
LLM hallucinations are not a bug — they are a property of the architecture. We show 5 techniques that significantly reduce them in production, and explain why eliminating them entirely is not possible.
August 29, 2025 · 9 min read
Invoices, purchase orders, and technical drawings fill industrial companies' inboxes every day — and most of them are still keyed in by hand. Document intelligence combining OCR and LLMs turns that into a structured, auditable data flow straight into ERP.
August 18, 2025 · 10 min read
"Looks good" is not a metric. How to build an eval set from real cases, where LLM-as-judge works and where it lies, and why regression tests before every release save more than code review.
July 31, 2025 · 9 min read
Most companies hire a data scientist and run into trouble six months later. The right AI team composition depends on what you are actually building — and whether building in-house makes sense at all.
July 18, 2025 · 11 min read
Most production incidents with AI agents don't stem from poor model reasoning — they come from poorly designed tools, malformed arguments, and missing retry logic. Here are the patterns that change that.
July 7, 2025 · 11 min read
A RAG pipeline without systematic evaluation is a black box — you don't know exactly where it fails. RAGAS separates retrieval from generation and gives you the metrics that actually matter.
June 20, 2025 · 10 min read
Most companies launch an AI project without a clear, measurable goal — and stall after the pilot. The first 90 days determine whether AI delivers real value or stays a PowerPoint presentation.
June 6, 2025 · 7 min read
A single agent handles 60% of use cases. Multi-agent handles 90%. The difference in cost, complexity, and maintenance is fivefold. The question is — are you in those 30% in between?
May 21, 2025 · 10 min read
When you need a model to truly understand your specialised documentation, SFT is not enough. Continued pretraining is a different tool with a different purpose — and in this article we explain when it makes sense and what it actually involves.
May 9, 2025 · 10 min read
Without tracing, an AI agent is a black box — you see what went in and what came out, but not why it made the decisions it made. Observability is what turns an agent into a system you can actually fix.
April 28, 2025 · 10 min read
"Return JSON" in the prompt works ninety percent of the time — and it's exactly that ten percent that breaks your production pipeline. We explain what lies behind it and how to achieve genuine reliability.
April 10, 2025 · 10 min read
Predictive maintenance with AI promises dozens of percent fewer outages — in practice, success depends almost entirely on data quality and history length. This article separates where ML genuinely helps from where a simpler rule does the job.
March 28, 2025 · 9 min read
If you're paying for the same system prompt thousands of times a day, you're paying unnecessarily. Prompt caching can cut input-token costs by 80–90 % — but only when you know what to cache and when it actually makes sense.
March 11, 2025 · 10 min read
A fine-tuned 8B model can outperform a generic 70B on a narrow domain — at a fraction of the cost and with no cloud dependency. When that holds true, and when the large base model wins instead.
February 27, 2025 · 10 min read
The embedding model is the cornerstone of every RAG system — and the decision most teams make once and never revisit. A practical selection framework for multilingual, on-prem, and cloud deployments.
February 14, 2025 · 10 min read
Sending company documents to a cloud LLM without a Data Processing Agreement? You are risking both a fine and your reputation. We show you exactly where data flows — and how to stop the leaks.
January 29, 2025 · 10 min read
Browser and computer-use agents promise to automate any website or application without an API — but real production reliability is far from the demo video. Here is an honest assessment.
January 17, 2025 · 10 min read
ReAct, Plan-and-Execute, or reflection — the choice of agent architecture determines reliability, cost, and debuggability. Here is a decision framework for real-world deployment.
December 30, 2024 · 9 min read
LLMs can automate 40–60 % of routine customer support queries — but without properly configured grounding and escalation they frustrate more customers than they save agent-hours.
December 18, 2024 · 10 min read
An AI agent without a well-designed memory layer is just an expensive chatbot. How context, vector memory, and summarisation work — and which layer each piece belongs to.
December 6, 2024 · 9 min read
MCP solves the M×N integration problem for AI agents — one open protocol instead of hundreds of custom connectors. What it means for your business in practice, and where the limits are.
November 19, 2024 · 10 min read
Quantisation cuts LLM VRAM requirements by 50–75 % with quality loss that typically stays below 5–8 % for 4-bit formats — the key is choosing the right format for your hardware and use-case.
November 7, 2024 · 9 min read
Most AI pilots look promising in the demo and quietly die before production. Not because the technology doesn't work — but because the pilot was never designed as a path to production.
October 21, 2024 · 10 min read
A RAG pipeline without citations is a black box for any organisation — you don't know what to trust. Grounding and attribution turn an AI answer into an auditable output that holds up under compliance review.
October 8, 2024 · 10 min read
A bad dataset doesn't just produce a weak model — it produces a model that confidently lies. This article covers how many examples you actually need for SFT and DPO, how to handle quality, deduplication, and data splits.
September 26, 2024 · 11 min read
Your first AI agent doesn't have to be complex — it has to be properly scoped. Here is the process that works in practice, including what to avoid at the start.
September 9, 2024 · 10 min read
A hospital, a bank, or a law firm cannot allow patient records, trade secrets, or privileged correspondence to pass through a third-party server. On-prem LLM is not merely a technical choice — it is an architectural commitment to the regulator.
August 28, 2024 · 10 min read
AI-powered defect detection achieves over 99% accuracy in production and processes hundreds of parts per minute — but only when the conditions most manufacturers underestimate are actually met.
August 12, 2024 · 10 min read
Red-teaming LLM applications uncovers jailbreaks, prompt injection, and data leakage before attackers find them in production — and shows where a single safeguard is not enough, but layered defence is.
July 29, 2024 · 6 min read
The GPU gets all the attention. The rest of the PC around it decides whether you'll utilize it at 95% or 60%. Five mistakes we see less often than a year ago — but less often still means they exist.
July 17, 2024 · 10 min read
Distillation transfers knowledge from a large teacher model into a small student — the result is a model that fits on an edge device or a low-cost inference server while retaining most of the original's capabilities.
July 1, 2024 · 10 min read
Choosing the right fine-tuning method is the difference between training on an ordinary server and needing a rented cluster. This article gives you a decision framework with concrete VRAM numbers for models from 7B to 70B.
June 18, 2024 · 10 min read
Classic RAG finds similar fragments — GraphRAG finds connections. If your questions require tracing relationships across documents, it's time to understand when the graph approach pays off and at what cost.
May 31, 2024 · 9 min read
The EU AI Act is not just a regulation for model developers — deploying companies have their own obligations, fines, and deadlines. Here is exactly what you need to have in order, and by when.
May 20, 2024 · 9 min read
Model merging combines multiple fine-tuned models into one without any additional training — in minutes instead of hours and without a GPU. When does it make sense, and when is it just blind experimentation?
May 7, 2024 · 10 min read
Token costs for a production agent can surprise you — loops, history, and repeated calls grow faster than you expect. Here is a framework for estimating costs realistically and keeping them under control.
April 19, 2024 · 10 min read
A million-token context window sounds compelling — but most production systems don't need it and pay dearly for every unnecessary token. When does a large context genuinely help, and when is intelligent RAG cheaper and more accurate?
April 8, 2024 · 9 min read
Chatbot, copilot, agent — three distinct levels of autonomy with different costs, risks, and payoffs. Before you sign a project, know what you are ordering.
March 22, 2024 · 10 min read
VRAM is the single parameter that definitively determines what your server can handle. How to calculate exactly how much you need — and where to save with quantization without a visible quality loss.
March 11, 2024 · 10 min read
Fine-tuning promises a model tailored to your domain — in practice, most projects fail before they ever reach production. Seven concrete causes and how to avoid them.
February 26, 2024 · 10 min read
Your choice of vector database will shape the performance, cost, and maintenance burden of your RAG system for years to come. A comparison of pgvector, Qdrant, Weaviate and Milvus with a decision framework based on scale, team size, and infrastructure.
February 9, 2024 · 10 min read
Prompt engineering is not magic — it is a discipline with proven patterns and clear limits. We show what works in practice, what is a myth, and why a good prompt will never replace an eval.
January 29, 2024 · 9 min read
Coding agents genuinely accelerate development — but not uniformly. Boilerplate, tests, and codebase exploration are their sweet spot; complex legacy refactoring and security-sensitive sections remain dangerous territory for blind trust.
January 11, 2024 · 10 min read
"We have data" does not mean "the data is ready for AI". Most companies go through an entire project and discover only at the end that their databases, files and logs are not in a state any model can meaningfully process. This article describes what an AI data pipeline looks like in practice.
December 29, 2023 · 11 min read
vLLM, SGLang and Ollama solve the same problem with three different approaches — picking the wrong tool can cost ten times the necessary infrastructure spend or deliver unacceptable latency in production.
December 18, 2023 · 10 min read
Most AI projects fail not because the technology does not work, but because nobody measured the baseline before starting. Here is a framework for calculating real ROI — including the costs that tend to get left out of presentations.
November 30, 2023 · 10 min read
Industrial PDFs are packed with tables, drawings and schematics that standard RAG simply skips. Multimodal RAG solves this — but the cost, complexity and the right architecture all depend on what your corpus actually contains.
November 17, 2023 · 10 min read
SFT, DPO, GRPO — each method solves a different problem. Knowing which to use and in what order saves weeks of work and protects your model from regression.
November 1, 2023 · 8 min read
95% of "let's deploy AI" conversations revolve around the model. 5% around the data. The real decision point is somewhere else entirely — who sees your sentences and how much you pay for them not to.
September 30, 2023 · 11 min read
4 months migrating an SK retailer with 12k SKUs from Shopify to Medusa.js. Three weeks lost to SK invoicing, two days to Algolia sync, and 0% of organic traffic died. The concrete playbook.
May 15, 2023 · 7 min read
Headless is a label. Underneath it are two architectural decisions that determine whether your investment will be profitable. If you confuse what you're buying, you get either too much or too little.
December 15, 2022 · 9 min read
Liquid cooling is currently the marketing slogan for AI infrastructure. Reality: for 60% of clients considering it, traditional precision air cooling would cover their need at 40% less cost. For the remaining 40%, it's the only physically possible solution.
August 5, 2022 · 11 min read
The NIS2 deadline is knocking and half of Slovak operators of essential services don't even have a Z/C model drawn, never mind implemented. Here's what it really costs to take IEC 62443-3-3 to SL-2 in an existing plant.
March 7, 2022 · 10 min read
The client asks "is it cheaper to build a new line or to refit the old one?" An answer of "it depends" is 100% correct and 100% unusable. You need five criteria, not a hunch.
October 6, 2021 · 11 min read
A 10 kWp panel on the roof without logic for self-consumption is 50% wasted money. PV + battery sizing doesn't start with panels, it starts with the daily consumption curve and what happens in a 3-day outage.
May 7, 2021 · 9 min read
The end of generous Zelená domácnostiam III subsidies pushed the solar PV break-even out by 18–30 months. Reality 2026: without battery 8–11 years, with battery 9–13 years. Household size matters more than panel brand.
December 7, 2020 · 9 min read
200 m² villa, three real quotes: KNX 38,000 EUR, Loxone 21,000 EUR, Home Assistant 9,500 EUR DIY. The difference isn't in comfort, but in who will service your BMS in 12 years.
July 29, 2020 · 7 min read
Tier III, Tier IV, Uptime Institute. Those are certificates. The SLA you actually receive from operations is determined by completely different things — and most of them can't be certified.
February 28, 2020 · 9 min read
In sub-1MW datacentres UPS is typically oversized by 60–80%. Five concrete sizing mistakes we've seen in real projects, plus numbers to verify before the tender starts.
September 30, 2019 · 9 min read
Clients ask "cobot or robot?" expecting a technology answer. The real answer hangs on two numbers — parts per hour and the cost of the fence — that nobody calculates in the first deck.
May 1, 2019 · 6 min read
The internet is full of comparison articles, KNX vs Z-Wave vs Zigbee vs Matter. The real decision doesn't depend on the protocol, but on the lifespan you expect. 5 years vs 25 years is a completely different build.
December 21, 2018 · 10 min read
KNX wins in residential, BACnet in commercial HVAC, LON stays in legacy industrial. The bus choice isn't a technology preference — it's a 20+ year maintenance commitment with a specific integrator ecosystem.
July 23, 2018 · 10 min read
A 12 EUR/m cable saving on a 220m campus link sounds like savings. In 8 years, when you need 100G and OM4 only stretches to 100m, it's an 80,000 EUR rebuild. The OS2 vs OM3/OM4/OM5 choice isn't made on today's CAPEX.
February 21, 2018 · 10 min read
The distributor shows the licence. That's 15% of the 5-year TCO. The rest sits in engineering hours, the database, client licences and modules you didn't think about at PO.
September 22, 2017 · 11 min read
The client booked a rolling mill decommissioning at 28 days. The reality was 42. Here's the day-by-day breakdown of what was planned, what broke, and how much that one cable somebody always cuts costs.
May 15, 2017 · 10 min read
The client asks "what bolt is good for UHV?" Wrong question. Right one: how much outgassing budget is left after the rest of the equipment.
December 14, 2016 · 8 min read
The client sends an inquiry: "We need a single-purpose machine for gluing." The first question we ask before quoting has nothing to do with technology — we calculate where the break-even sits versus a standard machine and a manual line.
July 15, 2016 · 9 min read
The client asks "is this equipment ATEX?" The answer is always "for which zone?" Without that question you buy more expensive than needed, or non-compliant and illegal.
February 15, 2016 · 9 min read
The client asks "we're migrating from S7 to ControlLogix, how long will it take?" A question nobody answers honestly before a code audit. And usually it isn't a technology choice but a boss choice.
October 7, 2015 · 9 min read
The client asks "why does our PLC sometimes lose communication?" In 80% of cases the PLC is innocent. The real fault sits in one of five places nobody checks.
May 8, 2015 · 8 min read
The client asks "what's better, OPC UA or Profinet?" — the answer is "both, but in different places". Modbus adds 30% complexity, OPC UA gives you interoperability, Profinet guarantees cycle time. None is universally better.
December 8, 2014 · 10 min read
A 22 kV substation from 1986 is still whispering. The DGA shows 380 ppm acetylene, oil BDV has dropped from 60 kV to 38 kV. The client asks — repair or buy new? Here's the decision grid that walks every part before PO.
July 9, 2014 · 10 min read
An ISO 50001 audit isn't a paper ceremony. With the right preparation, 5 days identifies 5–10% energy savings with payback under 3 years. With bad preparation the audit still passes, the client pays 8,000 EUR and a year later everything is the same.
February 28, 2014 · 7 min read
Clients often submit a project with the words "I want a new line". The line is the result. Before it lie six phases where it's decided whether OEE will be 60% or 92% after the first quarter.
September 30, 2013 · 7 min read
Half of the projects clients submit as "just a refurbishment" turn out, after first analysis, to actually be greenfield with more expensive conditions. The other half is submitted as greenfield, though brownfield would suffice. Here are the signals that decide.
April 30, 2013 · 6 min read
A quality switchgear cabinet isn't recognized by the brand of its components. It's recognized by whether a stranger electrician can open it after three years without a schematic in hand and know what to do.
November 30, 2012 · 9 min read
An inspection report with "no defects" is rare in practice. Across 200+ inspections in the last three years, we find on average 4.1 findings per installation — and 80% of them fall into the same twelve categories. If you run this list before the inspection, most can be fixed in one day.
July 1, 2012 · 6 min read
Twelve days of lost time during reinstallation would have paid for one decent notebook. The following notes come from 60-tonne machines, but they apply equally to a 60-kilogram switchgear cabinet.