Notes from practice
Blog
We write about what we work with ourselves: where document search breaks, how a knowledge graph differs from a vector database, when several agents do worse than one. No trend round-ups. Only mechanisms, limits and what we have checked by hand.
AI engineering
RAG system: what it is, when you need it and where it breaks
A RAG system makes a language model answer from your documents instead of from what it remembers of the internet. The idea is simple, and its failures are predictable. Almost all of them happen before the model: where the documents are cut into chunks, searched and passed to it.
Knowledge systems
Knowledge graph or vector search: which one fits your task
A knowledge graph and vector search answer the same request, “find what is relevant”, but they understand relevance differently. Vector search looks for what is similar in meaning; a graph looks for what is connected by structure. Mix them up, and a project spends a long time fixing the wrong place.
Agentic systems
When a multi-agent system loses to a single agent
A multi-agent system looks convincing on a slide: a whole team of specialists in place of a lone worker. In production it often turns out slower, more expensive and less predictable than one agent with good tools. This article is about where the line runs.
AI engineering
LLM hallucinations: the mechanism and the engineering measures against them
LLM hallucinations are not a model failure. They are the model’s normal mode of operation arriving at a result that is inconvenient for us. The mechanism matters more than any list of “prompts against hallucinations”: it shows which measures work and which only create a feeling of control.
Agentic systems
AI agent: what it is and what turns a system into an agent
Almost anything with a language model in it is called an AI agent today: from a script with a prompt to a system that runs a process on its own. The difference matters. It decides which tasks you can trust the system with and what control it needs.
AI and business
How to calculate the ROI of AI implementation so that the numbers survive scrutiny
Most calculations of the ROI of AI implementation fail one question from the finance director: what exactly will change in our costs. Below is a method that survives that question, and the mistakes that inflate the result several times over.
AI and business
Business process automation with AI: what to take on and what to leave alone
Business process automation with AI does not fit everywhere there is manual work. Many of the processes people tackle “with a neural network” are covered more cheaply and more reliably by ordinary rules. Below is how to tell one from the other while the budget is still unspent.
Metaintelligence architecture
Metaintelligence: why the environment around the model sets a system’s ceiling
We are used to measuring progress by the size of the model. But a system that carries work for months consists of more than a model. It has relationships between participants, limited resources, a way of reaching agreement and a trail from which any decision can be taken apart. This series is about how such a layer is built and why it sets the ceiling of what is possible more firmly than the next generation of weights.
Metaintelligence architecture
The agent as a bounded local world
In most systems an agent is built as a function with a prompt. Such an agent cannot refuse work, come to an end or run out of resources. Yet it is being bounded that makes a participant fit for an environment where nobody controls everyone.
Metaintelligence architecture
A link as a first-class object: a relationship with state and history
In most systems the link between participants is invisible: there are calls, a queue, a shared channel. Make the relationship an object in its own right, with its own state and history, and things appear that do not exist otherwise: participants can reach an agreement, revise its terms and part cleanly.
Metaintelligence architecture
Living graph: topology shaped by work
An architecture diagram goes out of date the moment it is drawn. A living graph tries to close the gap between the diagram and the system: links here are not designed in advance. They arise from the participants’ local decisions and disappear when they stop being useful.
Metaintelligence architecture
Consensus without a centre: deciding together without taking anyone’s word for it
As long as a system has a master node, the question of a shared decision does not arise: the master decides, the rest carry it out. As soon as there is no master, or it cannot be trusted unconditionally, a problem appears that ordinary agent setups simply do not deal with.
Metaintelligence architecture
Agent economics: autonomy needs a budget
The most reliable limit on behaviour is a resource that runs out, not an instruction in the prompt. When every action and every thought is debited from an account, “don’t do anything unnecessary” stops being a request and becomes a law of the environment.
Want your problem looked at?
Describe what is not working or what you want to build. We will answer on substance, and if the problem has a simpler solution than it seems, we will say so.
AVA calculator: your figures, no form, no price from us.