Who are you?
We are mindos: a team of senior people who have led companies and their technology at C-level, in markets around the world.
We make your enterprise autonomous.
The enterprise of the future is already running.
It runs on swarms of agents that plan and act, verifying one another's work, and takes more decisions in a day than any organisation before it. Leading it requires an understanding of what a swarm can do and where it fails.
Most of what is sold as an agent today is a tool for a single task. Taking an enterprise to that scale is a different problem, one of verification and of the decisions around it, which no purchase settles.
The nearest analogue to this change is fire. It burns those who ignore it and those who rush at it; the ones who learned to handle it built everything that came after. Staying relevant means learning to handle it.
What would you like to know?
- Who are you?
- What do you do?
- What have you built?
- Who have you worked with?
- How much experience do you have?
- How old is mindos?
- How would we start?
- What am I looking at?
What do you do?
We help you turn your enterprise into a consolidated, autonomous whole. Agents carry the operation: they plan, act and check one another's work. Your people set the direction, and the judgement remains theirs. We build those systems ourselves, and we advise the leaders who have to steer them.
What have you built?
That is better seen than described. Let's meet, and we will be happy to show you our most recent work.
Who have you worked with?
Over our careers, with clients from the United States to Australia, from China and Japan to Germany and Hungary. Some were startups still in stealth; some served hundreds of millions of users. Part of that work had never been done before, and part of it shows in a country's GDP. We name none of them here, and we would not name you.
How much experience do you have?
In every field we advise on, someone here has more than twenty-five years behind them, AI included. At a large consultancy, a senior is five years out of university. Our people have run companies, sold some and lost some, led others through a crisis, held leadership posts in critical infrastructure, and built technology that did not exist before. That is the judgement you hire.
How old is mindos?
mindos came together in 2025. The stories behind it are older: the earliest goes back more than forty years.
How would we start?
With a conversation, and then a close look at one area of your operation: where decisions wait, where work is copied from one system into another, where a mistake costs the most. From that comes a plan, and a first system small enough to prove itself before the larger ones follow. Write to start@mindos.hu.
What am I looking at?
Several ideas, drawn as one picture. It is a swarm of agents at work: every light is an agent on its errand. A building that turns blue sends them out, and a white light checks their work. It is also a company: the buildings are its teams, and the work moves between them. The towers keep what the city has learned. Agents and company alike work the way a city does. The blue light ahead of you is carrying your question.
- Reasoning interleaved with acting
- Search over reasoning states
- Verification and failure modes
- Feedback loops
- Collaboration with assigned roles
- Dense retrieval, late interaction
- Generation from retrieved evidence
- Position in a long context
- Hierarchical and graph indices
- Measured citation
- Streaming recognition
- Pretrained speech models
- Neural codec language models
- Turn-taking
- Multi-horizon forecasting
- Pretrained forecasters
- Learned weather
- Grid operation as sequential decisions
- Calibrated uncertainty
- Scale and its economics
- Low-rank adaptation
- Preference and reasoning
- Latency
- Bias in the judge
- Sequence
- Build or buy
- Measure before scaling
Reasoning interleaved with acting
- The idea
- A model that can only talk has to guess. One that can look things up works like a good dispatcher, who thinks, checks and thinks again with the result in hand.
- How we use it
- Our diagnostic agent doesn't guess at a fault: before it names a cause it reads the device's own log at the maker's monitoring service, through a short list of read-only lookups. Add a call that writes, and the program refuses to start. On the phone, our agents answer from your documents, fetched as the question is asked.
- What it solves for you
- A cause is named from the device's own record instead of a plausible guess. On a call, every search the agent makes is kept, so a disputed case can be retraced.
Search over reasoning states
- The idea
- A model usually commits to its first line of thought. It can instead draft several, score each and back out of the dead ends before it answers.
- How we use it
- Our fault diagnosis takes its time over a device log: it names the likeliest cause, the rivals still standing and why each is less likely, with your experts' rules open before it. The heavier version, second opinions with a tie-breaker, we built and measured; it did not beat the simpler way, so it stayed out.
- What it solves for you
- Hard cases stop getting the first answer that came to mind. Your engineer sees which alternatives were rejected and why, and can overrule the system without an investigation.
Verification and failure modes
- The idea
- An auditor who checks only the bottom line approves accounts in which two errors cancel out. Agents need every step examined, by something other than themselves.
- How we use it
- No agent of ours signs off its own work. Code checks every value an agent records, a separate marker reads its answers, and a serial number is looked up in the maker's register before a diagnosis rests on it. Before anything is merged, reviewing agents try to break it, and a second set must confirm each fault they claim.
- What it solves for you
- A wrong value is stopped where it is recorded, not found weeks later in a report. When something does fail, the log names the step, and repair starts there.
Feedback loops
- The idea
- An agent that fails can note why and reread the note before its next attempt. The principle is a technician's logbook: what broke, what fixed it.
- How we use it
- A lesson counts only when firm evidence stands behind it, such as an engineer closing the fault and the code not coming back. A person approves any note that would change the agent's behaviour.
- What it solves for you
- An error corrected once is unlikely to return. The system improves in service without a new project each time, and your people decide what it may learn.
Collaboration with assigned roles
- The idea
- Nobody asks one employee to sell, check the contract and approve the payment. Agents, too, often work better in roles, one drafting, another challenging the draft.
- How we use it
- We split a process into roles and enforce each agent's procedure in software, so no caller can talk an agent past a step. Each agent sees only the data its role needs.
- What it solves for you
- A new approval limit or an extra check means editing one role and leaving the rest alone. Roles are tested and audited separately, and no single agent holds every key.
Dense retrieval, late interaction
- The idea
- Keyword search finds the words you typed. Search by meaning matches a customer who says the unit has gone dead to the manual page on power supply faults.
- How we use it
- We index your manuals passage by passage, each with its document and page, and your closed cases one by one. Every question is searched by its words and by its meaning at once, and which meaning model serves you is decided by measurement, on questions put the way your engineers and customers put them.
- What it solves for you
- A new colleague finds the right page without knowing the official term for the problem. Fewer questions land on the one engineer who knows the manuals by heart.
Generation from retrieved evidence
- The idea
- A model's own memory is fluent and sometimes wrong. An open-book exam works better: fetch the relevant pages first, then answer with them open on the desk.
- How we use it
- Our agents answer from your approved documents, fetched at the moment of the question. Where the documents are silent, the agent says so and passes the question to a person.
- What it solves for you
- Answers follow your current tariff sheet and procedures, not the internet's average. When a rule changes, you replace the document and the answers change with it, without retraining.
Position in a long context
- The idea
- Hand a model fifty pages and it reads like a tired person. The opening and the ending register; a fact buried in the middle is easily missed.
- How we use it
- We give the model few passages, only the best handful. Each is cut to a paragraph and the whole kept under a fixed size. When something must go, it is the weakest, and the test search shows the exact list the agent saw.
- What it solves for you
- This discipline guards against a quiet failure, in which the right document went in and the wrong answer still came out. Shorter inputs also cost less, since every page read is billed.
Hierarchical and graph indices
- The idea
- A question such as which complaints recur with one device type has no single page that answers it. It needs an overview, like a book's chapter summaries and index.
- How we use it
- Over the whole fleet we keep a register that links site, device, software version and fault code. A code that appears on many devices at once is flagged, since that usually has one outside cause.
- What it solves for you
- Pattern questions draw on the whole archive, not on the few documents a search turned up. A failing component batch can show up while the fix is still cheap.
Measured citation
- The idea
- A footnote is worth something only if the page it points to supports the sentence. Such support can be counted: how many statements are borne out by their sources, and how many footnotes were needed.
- How we use it
- Our knowledge search shows document and page under every answer, and the call's page keeps the list the agent saw. Before launch we measure, on questions in your own language, how often the right passage reaches the agent.
- What it solves for you
- When a regulator asks where a figure came from, the answer is a document and a page. How often the right passage is found is a measured figure, not a promise.
Streaming recognition
- The idea
- Many speech recognisers write nothing until you have stopped talking. A streaming one writes while you speak, like a stenographer, and judges for itself when you have finished.
- How we use it
- The transcript grows word by word, but we act only on the finished sentence, because the half-heard one is too often wrong. Before the caller says a word, the dispatcher has looked up their number and any call of ours that is waiting for them.
- What it solves for you
- A known caller is recognised before they explain who they are. A sentence split by a poor line is read as one turn, not answered piece by piece.
Pretrained speech models
- The idea
- Today's speech recognisers have heard more speech than a person could in many lifetimes, much of it mumbled or spoken over traffic. They don't start from zero with your first caller.
- How we use it
- We test the recogniser in your customers' language on the telephone's own narrow band, with scripted callers, before a real call is taken. Your vocabulary we handle after recognition: a list of your product and type names turns every spelling the recogniser produces into the form your systems use.
- What it solves for you
- The recogniser may spell one type name five ways over a weak line; your systems receive the one form they know. A spelling the list has never met becomes a proposal for a person to approve, not a silent error.
Neural codec language models
- The idea
- Older synthetic voices were stitched together from recorded snippets. The newest are generated sliver by sliver, each sound predicted from what came before, so intonation and emphasis follow the meaning.
- How we use it
- We keep text in the middle. The agent hears, thinks in writing and then speaks, so every word it said can be read and checked afterwards. Before a sentence is spoken, our own code writes each amount, date and order number out the way a person would say it.
- What it solves for you
- A robotic voice sends callers straight to the human queue, and the business case goes with them. A natural one keeps them on the line, and a clearly spoken amount is rarely misheard.
Turn-taking
- The idea
- People don't wait for silence before answering. They hear from its rhythm and pitch that a sentence is ending, and a model can learn that timing from recorded conversations.
- How we use it
- Our voice agents judge that the caller has finished from the sound of the line together with the recogniser's finished sentence, on thresholds we measure in test calls. When the caller cuts in, the agent falls silent and listens; only the notice the law requires is spoken to the end.
- What it solves for you
- The agent rarely talks over your customer or leaves them wondering whether the line went dead. Either fault is enough to make a caller ask for a person.
Multi-horizon forecasting
- The idea
- A useful forecast covers every hour of the coming days at once. The model keeps what is known ahead, such as holidays and the weather forecast, apart from what only the meter can tell afterwards.
- How we use it
- A forecast from us will take the two kinds of input separately, for generation and for demand alike, and will show which inputs weighed most.
- What it solves for you
- Your scheduler works from one model for intraday and day-ahead decisions, where separate ones can disagree at the joins. When a forecast misses, you can see which input it leaned on.
Pretrained forecasters
- The idea
- Some forecasting models have studied so many histories, from road traffic to power demand, that they can project one they have never met, as a seasoned analyst reads an unfamiliar chart.
- How we use it
- On a new site we would start with one of these, on the first weeks of data. Any custom model we propose must then outdo it on your site's history, or we don't build it.
- What it solves for you
- You never pay for bespoke modelling that a general forecaster would have matched. A newly commissioned plant gets a usable forecast within weeks, without a year of history.
Learned weather
- The idea
- Classical forecasts compute the atmosphere's physics on a supercomputer, over hours. The newer ones learned from decades of recorded weather; they run in minutes and do better on most measures.
- How we use it
- In a forecast of ours the learned weather services would go in beside the classical ones, never in their place. Each source would keep its place only by being scored against what your site measured.
- What it solves for you
- Weather causes the largest error in most renewable schedules, and the imbalance bill follows it. A better input, refreshed more often, cuts that error at its source.
Grid operation as sequential decisions
- The idea
- Running a network is a chain of decisions, each switch changing the options for the next. An agent can practise on a simulated grid, where a blackout costs nothing.
- How we use it
- From your network model we would first build the simulator, a copy that refuses forbidden switching. What the agent learned would reach the control room as a suggestion, as our fault diagnosis does today: it proposes, a person decides.
- What it solves for you
- Congestion can often be relieved by reconfiguring the network instead of paying for redispatch or curtailment. Operators get a tested second opinion, and nothing untried touches the live grid.
Calibrated uncertainty
- The idea
- A forecast range should keep its word: if it claims to hold the real value nine times in ten, it must. The method counts the misses and corrects the range, as a thermostat would.
- How we use it
- No forecast will leave us without such a range, and we will keep counting how often it holds. It is a rule we already keep elsewhere: our diagnoses state how strong their evidence is, and our estimates give a likeliest figure between a best and a worst case.
- What it solves for you
- Reserves, bids and hedges are sized to a measured risk, not a gut feeling. A range that goes on widening also warns that the model has gone stale, before losses do.
Scale and its economics
- The idea
- Researchers have mapped how model size, training data and cost trade off. The practical lesson: a smaller model, well trained, often matches a large one for far less.
- How we use it
- We choose the model per step, not per project. A small, fast one pulls a date out of a sentence or marks an answer; the large one handles the rare hard judgement. What plain rules can decide, such as where a call belongs, gets no model at all.
- What it solves for you
- Running cost grows with every call, so this choice decides whether the project pays.
Low-rank adaptation
- The idea
- Teaching a general model your trade once meant retraining all of it. Now the model stays frozen and a small add-on is trained on top, as an experienced recruit picks up your house style.
- How we use it
- On the shorthand of fault reports we have so far left the model untouched and let it read similar closed cases instead, and we measured the gain on cases it had never seen. We would train an add-on only where that falls short, from your own examples, on a single machine.
- What it solves for you
- A specialist model costs a workstation, not a data centre. The training data can stay inside your walls, and the add-on is a small file you own.
Preference and reasoning
- The idea
- Describing a good answer is hard; picking the better of two is easy. Models are shaped by such choices or by written principles, and learn to reason where answers can be checked.
- How we use it
- Your house rules are written down as a plain document that stands before every step the agent takes, and your staff's own verdicts serve as the test it must keep passing. Where tone matters more, your best people would choose between pairs of draft replies, and their choices would steer it.
- What it solves for you
- The agent behaves more as your company would, down to its caution about promises. Its principles are a plain document your compliance team can approve.
Latency
- The idea
- On the phone, a pause much over a second sounds like a fault. A small model can draft ahead while a large one approves or corrects, and the same answer arrives sooner.
- How we use it
- Drafting ahead is the model maker's business. Ours is the clock: every turn of a call has a time budget, and every stage inside it is timed, from the end of the caller's sentence to the first sound of the reply.
- What it solves for you
- Callers who meet a long pause repeat themselves or hang up, and both cost you. A slow call shows up as one slow stage, which can be replaced without touching the rest.
Bias in the judge
- The idea
- A test run produces more conversations than any team will read, so a model marks them. That marker has habits: it favours the first answer, the longer one, and prose like its own.
- How we use it
- We hold the marker to two yardsticks: the verdicts your staff gave by hand, and panels of independent markers that never saw its answer. We rewrite its instructions until it agrees with those panels as well as they agree with each other, and where it still proves lenient, the report says so.
- What it solves for you
- The quality figure on your dashboard means what it says, and no rollout rests on a flattered score. A supplier's benchmark can be questioned the same way.
Sequence
- The idea
- Where to start is a business decision before a technical one. A first process too rare, too tangled or too short of data puts people off the rest.
- How we use it
- We read what the operation has written down and ask its people for what is missing. Each candidate is scored on how complete its brief is, how clear its scope, how risky the technique and whether its data is reachable. Where access is missing we say so and give no figure.
- What it solves for you
- You avoid the pilot that spends its budget waiting for access rights. The board gets a ranked list with reasons, its first item chosen to fund the second.
Build or buy
- The idea
- Some parts you must own, because they hold your data or set you apart; the rest you can rent. The purchase price is rarely the biggest figure.
- How we use it
- We price each option over the years it will run, at full volume, with staff and exit counted in, and we say plainly what a figure leaves out. Where buying wins we say so, though building is our trade: in our own voice agent the recogniser, the voice and the line are rented; the flow engine and the knowledge store are ours.
- What it solves for you
- The costly surprises surface on paper first: a subscription cheap in the pilot and ruinous at scale, or an archive of recordings stuck in a supplier's format.
Measure before scaling
- The idea
- A pilot is an experiment, and an experiment needs its yardstick before it begins. Set afterwards, the yardstick bends, and any result can be argued either way.
- How we use it
- Before anything is switched on we measure how the process performs today. Our fault diagnosis began that way, with the service team's own closed cases as the yardstick and the old way measured against it first. Together we write down which result means expand, which adjust, and which stop.
- What it solves for you
- The pilot can't drift for months while people argue over what success was. The decision to scale becomes a reading taken from a gauge everyone accepted beforehand.
We have gone this way ourselves, and we can light the dark ahead of you. We build the systems, and we advise on the leadership and the technology the change demands.