Custom AI systemsAccountable agentsKnowledge graphsData engineeringPrivate modelsAutomationMarket and financial analysisCustom AI systemsAccountable agentsKnowledge graphsData engineeringPrivate modelsAutomationMarket and financial analysis
Every system we build starts with how your business already works. From there it goes down, layer by layer, to the machines it runs on. Pick a business and follow one piece of work all the way through.
The task
Layer 00
Your business
We start with you, not with technology: three days inside how you work, who does what, where the time goes and which rules are never broken.
Your task, here
What we design
A map of the work as it really happens
The rules that must never be broken
The numbers that will prove it worked
How it fits your style
Your words, your approvals, your tools. If your team says "job" instead of "ticket", so does the system.
Passes downGoals, rules and deadlinesComes back upResults, in your terms
Layer 01
Workflows
The work becomes flows: what starts each one, the steps, who owns each step and what done means.
Your task, here
What we design
Triggers and steps
An owner for every step, person or agent
A written definition of done
How it fits your style
We automate the steps your team dislikes and keep people on the ones where judgement matters.
Passes downTasks, each with an owner and a definition of doneComes back upFinished work, ready to check
Layer 02
Agents
Each step is handled by a specialist agent. Not one big chatbot, but a small team, each with a written job.
Your task, here
02.1
Behaviour spec
A standing rulebook, not a one-off instruction: what the agent reads first, what it may touch, what counts as done and when it must ask you. Every agent is tested against its rulebook.
02.2
A soul for each role
Each agent has a written working style. The reviewer is sceptical, the drafter is brief, the tester trusts nothing it has not run itself. These are rules we can check, not personalities for show.
02.3
Trust is earned
Every agent starts on probation, suggesting rather than doing. It moves up one rung at a time, for one kind of action at a time, when its record earns it and you agree.
02.4
Evaluation
One good run is luck, so we run the same jobs many times. Agents are scored on cases taken from real past mistakes, including cases held back that they have never seen, by checks that never rely on the agent's own word.
02.5
Darwinian improvement
Agents improve by selection. A changed version competes with the current one in a sealed room, and wins the job only if it loses no case and wins at least one. A person approves the switch, and every version is kept.
Researcher
Working style
Curious, cites everything
Reads first
Your records, then the web
Never
States a fact without a source
Asks you when
Two sources disagree
Reviewer
Working style
Sceptical, evidence first
Reads first
The brief, before the work
Never
Takes the author's word for it
Asks you when
A risk needs your decision
Tester
Working style
Trusts only what it ran
Reads first
The spec, never the code
Never
Passes its own work
Asks you when
A check cannot be run
How hands-on are you?
5Runs: you watch the dashboard
4Acts: you review it weekly
3Acts: you are told
2Drafts: you approve
1Suggests: you do it
×Never by an agentMakes a payment
Reads your dataDrafts a documentEmails a client
Payments stay with a person at every setting.
Illustration
v1Current version, doing the job
v2Lost one case. Kept on file, not promoted.
v3Won two cases, lost none, held up on unseen ones
✓You approve: v3 takes over, v1 is kept
What we design
A rulebook and a soul for every role
A trust level for every kind of action
Test cases from your own business
How it fits your style
You choose how hands-on you are. Pick an answer above and the ladder moves.
Passes downProposed actionsComes back upWork, with its evidence
Layer 03
Harness
Agents run inside a harness: the frame that decides what they may do, checks what they did and decides when work is finished.
Your task, here
Four kinds of action
AutomaticReading a report
Checked firstUpdating a draft
Needs a personSending to a client
NeverDeleting records
01The agent says it is done
02Checks: 2 of 14 failing, sent back
03The agent fixes it
04Checks: 14 of 14 passing
05Reviewer: no open objections
06You approve, and it ships
What we design
Done is decided by checks, not by the agent: the tests pass and no objection is left open, or the work goes back.
Blind review: the checker writes down what it expects before it sees the work, so no one marks their own homework.
Handover notes, so any agent or person can pick the work up exactly where it stopped.
Sandboxes for special projects: a sealed room with a copy of the data and one supervised door out.
How it fits your style
Your approval points are written into the harness, not left to anyone's memory.
Passes downApproved stepsComes back upChecked results
Layer 04
Context
An agent is only as good as what it sees. Context engineering decides what goes into each agent's view, and what stays out.
Your task, here
One agent's view, for one step
Your rulesThe taskFacts fetched for itNotes from last time
Left outThe rest of the databaseOther clientsOld drafts
What we design
Just in time: agents keep an index and fetch what they need, instead of carrying everything.
Clean desks: helpers work in their own small space and report back in a paragraph.
Notes between shifts: long jobs leave notes, so the next session starts where the last one ended.
Prompts as specs: written, reviewed, versioned and tested like code, never improvised.
Reuse: when an agent re-reads the same briefing within minutes, it costs about a tenth of the first read.
How it fits your style
Your rules and your vocabulary sit at the top of every agent's view.
Passes downRequests for exactly what is neededComes back upFacts, each with its source
Layer 05
Data
Data sits on several shelves, each with its own cost, speed and caching. The right fact lives on the right shelf.
Your task, here
SpeedCost to use
DeskWhat an agent is working on right now. Instant, and the most expensive per word.
Reuse shelfBriefings reused within minutes cost about a tenth to read again.
RecordsTyped records: clients, invoices, properties. The source of truth.
SearchFinds meaning, not only words, across documents and notes.
GraphHow things connect, and how that changed over time.
ArchiveEverything, kept cheaply. Slower to reach, never lost.
Relative scale: more dots, faster or costlier.
The shared memory
Brain OS
Our name for the memory every agent reads before it acts: records, notes, decisions and what was learned last time. Written summaries are rebuilt from the records, so they never drift from the facts. When the two disagree, the record wins.
What we design
Which shelf each kind of fact lives on
One source of truth, with summaries rebuilt from it
Nothing deleted: old records are closed, never erased
How it fits your style
Where we can, your existing systems stay the records shelf. We connect to them instead of moving everything.
Passes downFacts and how they connectComes back upAnswers, with dates and sources
Layer 06
Graph
Facts are joined into a graph of people, companies, properties and contracts, and how they connect. Every connection carries two dates.
Your task, here
Owner of Apartment 12
In the world
Arta K.Besnik M.
In our records
Arta K.Besnik M.
True on that date, as we know it now
What we believed on that date
What we design
Two clocks: when something was true in the world, and when we learned it.
Corrections without erasing: an old fact is closed with a date, never deleted.
A source on every fact: the email, invoice or call it came from.
One person, many spellings: records matched into one, with a note of how, and reversible.
How it fits your style
Audits and disputes get simple: pick a date and see exactly what was known then.
Passes downRequests to your systemsComes back upData, checked at the door
Layer 07
Connections
Agents reach your tools through connectors built on the Model Context Protocol (MCP), an open standard. One connector per system, each with only the access its job needs.
Your task, here
AgentScopedMCP connectorLoggedYour system
The keys stay with the connector. The agent only asks for a job to be done.
What we design
One connector per system: accounting, email, files, calendar.
Least access: read-only unless writing is the job, narrow scopes, short-lived permissions.
Keys come from a vault at the moment of use and never reach the model.
Every call is logged: who asked, for what, and what came back.
Outside text is data, never instructions, so a hostile email cannot steer an agent.
How it fits your style
We connect to the tools you already use. Your team learns no new software.
Passes downWork to runComes back upResults, fingerprinted
Layer 08
Compute
Work runs where it fits: local models for private and bulk work, frontier models only for the hardest reasoning.
Your task, here
What leaves the building
Figures the reasoning needs stay. Names and accounts become codes, and the answer is turned back into real names here.
Every decision, chained
Requesta3f9 0c12
Answer7c1e 44d0← a3f9
Your approval0b52 e9a7← 7c1e
Each record carries the fingerprint of the one before it. Change any line and every later link breaks, and the seal is kept outside our own systems.
What we design
Local first: sorting, summarising and redaction run on machines we control.
Frontier when needed: only the hard reasoning goes out, never all the data.
Pseudonymised before sending: names, accounts and addresses become codes. The frontier model reasons over codes, and real values are restored locally.
Tamper-evident decisions: every request, answer and approval is fingerprinted and chained to the one before.
Sealed rooms: special projects run in isolated sandboxes with their own copy of the data.
How it fits your style
You decide what may leave the building. Some clients keep everything local.
Layer 09
Security
Security is not a layer added at the end. It runs through every layer above.
Private dataUntrusted contentSends messages out
Never all three.
An AI becomes dangerous when it can see private data, read untrusted content and send messages out, all at once. We design every agent so it never holds all three.
Your business
Your rules are written down, including what must never happen.
Agents
Each agent touches only what its job needs.
Harness
Risky actions wait for a person. Money never moves without one.
Context
Outside text is treated as data, never as instructions.
Data
Sensitive records stay local. Nothing is deleted, and every change is dated.
Connections
Scoped, short-lived access. Keys never reach a model.
Compute
Pseudonymised before it leaves. Every decision chained.
Between clients
We share methods between clients, never information.
People
An accountable engineer you can call, ready to step in at any moment.
How it fits your style
We design for Albanian and EU privacy law from the first day, and tell you where every piece of your data lives.
Layer 10
Languages
We choose each language for the job it does best.
Python
Data, models and agents. The deepest toolkit for AI and data work.
TypeScript
Interfaces and connectors. Typed, quick to build, and it runs in every browser.
Rust
Fast, memory-safe services where speed and reliability matter most.
SQL
Records and their history. Still the clearest way to ask a database a question.
Shell
The glue: schedules, scripts and automation on our own machines.
Plain text
Records and notes in open formats, so people and agents read the same files.
How it fits your style
If your team already works in a language, we build in it where it fits.