A deterministic core, a model only at the edges
Pricing, policy, classification and reconciliation are code. The model parses language and explains results. It never decides something that has a right answer.
BUILT, NOT PITCHED
Six systems.
Each one already runs.
Six problems a business already pays people to solve, built as working software rather than slideware. Tax reconciliation, agent governance, EU AI Act evidence, agentic checkout, nonprofit marketing and retail stock and cost. Every number on this page is produced by the system it describes.
STAGE: COMPLETE AND RUNNING, NO CUSTOMER YET. EACH ONE CAN GO TO A FIRST PILOT AS IT STANDS.
01
A marketing coordinator
for every nonprofit.
Social media assistant
THE PROBLEM
Nonprofit marketing is one person, ten channels and no research time. Generic AI copy fails here for one reason: it is unsourced, so nothing it claims can be published without checking it first.
WHAT IT DOES
It researches the organisation before it writes a word, drafts per platform from an editable rule file, and remembers corrections. Tell it once that you are warm and grassroots rather than corporate, and every later draft holds that, across sessions and restarts.
Sources on every claim, so nothing publishable is unattributed. A new platform is a text file, not a deployment.
WHO WOULD BUY IT: agencies and nonprofit networks, per organisation, per month
02
The books against what
the state actually received.
myDATA reconciliation
THE PROBLEM
Every Greek business must match its books to the tax office by hand. When the two disagree, somebody finds it document by document, usually late, and the penalty lands on the client.
WHAT IT DOES
It pairs documents on identity, then on a tolerant window so a one-day or one-cent difference is still the same invoice, classifies ten discrepancy types by rule, and explains each one in Greek. It proposes the correction and never files it: a dry run until a person approves.
Ten discrepancy types, classified by rule rather than by a model, so the same period gives the identical answer every run.
WHO WOULD BUY IT: Greek accounting firms, per client company, per month
03
A policy in the request path,
before the tool runs.
Agent control plane
THE PROBLEM
The moment an agent can call tools it can delete, spend, email and leak. Most teams ship that with a prompt asking it to behave. What is missing is the layer that decides, before the call runs, whether it is allowed, and leaves a record either way.
WHAT IT DOES
Allow, deny or route to a human, decided first. Budgets checked before the upstream, so a refused call cannot exhaust a tenant. Injection scanning in both directions. Everything traced, refusals included, because a gateway that logs only what it allowed cannot answer an auditor.
25 of 25 attacks caught with 0 of 25 false positives, and an unmatched tool denied by default so adding a tool never silently adds a capability.
WHO WOULD BUY IT: any company putting agents near production, per seat or per agent
04
Compliance packs,
retrieved rather than recalled.
EU AI Act conformity
THE PROBLEM
The Act is in force and being confidently wrong is the liability. A model asked to write an evidence file from memory produces a fluent document citing articles that do not say what it claims, which is worse than no document at all.
WHAT IT DOES
The model never recalls an article, it resolves one from a pinned, versioned corpus, and a citation that fails to resolve raises rather than finding the nearest match. Closed lists are settled in code at zero cost. A claim survives only if a separate verifier agrees its evidence supports it.
Zero unsourced claims as a structural gate rather than a target, and an edit made after issue is caught by the content fingerprint.
WHO WOULD BUY IT: compliance consultancies and EU deployers, per assessment, then a retainer
05
A merchant that agents
can buy from, safely.
Agent commerce
THE PROBLEM
Agentic checkout is arriving through three incompatible protocols and none has won. A merchant that picks one bets the integration on an unsettled outcome, and one that lets an autonomous buyer negotiate without structural limits gets exactly the sale it deserves.
WHAT IT DOES
One interface behind all three protocols. The margin floor is arithmetic in integer minor units, so no discount at any customer class can price below it. Every mandate is verified before money moves, and nonces, holds and settlements outlive the process.
Oversell, double charge and replay all refused, including across a restart, with no model call anywhere on the money path.
WHO WOULD BUY IT: merchants and platforms selling to autonomous buyers, per transaction or licensed
06
Stock, cost and orders
on one database.
Retail ERP suite
THE PROBLEM
A retailer's real numbers live in six tools that disagree. Nobody can answer what a unit actually cost or what the month actually made, because no single system holds both halves.
WHAT IT DOES
Freight and duty are allocated across a shipment by value, so price is set on landed cost rather than on the invoice. Stock is never stored, it is derived from an append-only movement ledger, so every unit traces back to the reception, sale or transfer that produced it.
On one shipment the landed cost ran 8.9% above the supplier price, which is exactly the margin a retailer loses by pricing off the invoice.
WHO WOULD BUY IT: independent retailers and small chains, per location, per month
HOW THEY ARE BUILT
The same four rules,
in all six.
The hard part is not the model. It is everything around it: what the model is allowed to decide, what holds when it is wrong, and whether anyone can check afterwards.
Structural defences, not instructions
The margin floor, the replay nonce, the tenant scope and the durable ledgers are code that a prompt cannot talk its way around. They hold when detection fails.
Measured, not asserted
Every figure on these screens is produced by the system's own code and re-checked by its test suite. The suites redden under mutation, so green is evidence rather than decoration.
Extended by you, not by me
Capabilities are files, not deployments. A non-engineer adds a platform, a skill or a research source and it is live on the next request.
Want one of these
pointed at your data?
Each of these can go to a first pilot as it stands: the interfaces are real, the rules are enforced in code, and every figure comes from the running system. The next step for any of them is a pilot against live data.