Evatech

Among the cleanest centrality cases on the whole Chartis roster, and the opposite pole from every platform vendor on it. The stated product is proprietary machine learning models that quantify operational business metrics to derive the true revenue and profit of a small business, explicitly in place of reading its financial statements.

Apply the Oxane refinement and the answer is immediate: strip the models and there is nothing left for the software to operate on, because the entire proposition is inferring a figure the borrower has not supplied. Notable in context: twelve large vendors from this roster graded C on this axis and the smallest name on it is the one where the model is the product.

Asserts an outcome and names no control, and the outcome asserted is the removal of the decision interval itself. The headline promise is approving small business loans in seconds with no financials needed. Nothing states whether a credit officer sees the result, whether any score band routes to manual review, what confidence attaches to a derived revenue figure, or what happens when the model's estimate and a borrower's own account of its business disagree.

The standing lending rule applies with full force and is unanswered: in a lending workflow the automated decline is where oversight matters most and is routinely the one point left undescribed. Ask about the knockout specifically.

No validation methodology, no accuracy figure, no backtesting, no discriminatory power measure, no drift monitoring, no versioning, no external assessment, for a firm whose only product is a model. One published number deserves comment because it is the right kind of metric presented in an unusable way: under 1 percent non performing loans on unsecured small business lending is a portfolio outcome and arguably the most meaningful evidence a credit model can offer, and it arrives with no institution, no portfolio size, no observation period, no vintage and no comparison against the lender's prior book.

A performance claim without a denominator is not model transparency. Contrast with Loxon on the same roster, which never markets on explainability and publishes backtesting, significance testing and reject inference.

The ComplyCube shape precisely: institutions named, numbers published, and the two halves never joined. Eight bank logos appear under a clients heading, including three large European banking groups, which is a genuinely strong list for a firm this size.

Beside them sit four aggregate figures with no institution attached to any of them: more than 100 percent return per project, under 1 percent non performing loans on unsecured small business lending, more than 50 projects and 830 million dollars of additional client income. No case study, no named executive, no quote, no attributed result anywhere. Held at B because the names are real and stated as clients; the standing mixed logo wall caution is recorded, since nothing distinguishes a customer from a pilot or a partner.

Nothing published on training data, pooling, retention or customer data use. The question is sharper than usual because the models are described as proprietary and the method depends on learning relationships between observable business activity and real turnover, which is knowledge accumulated across borrowers. Whether one bank's portfolio outcomes improve the model another bank buys is exactly the pooled corpus question and it is neither claimed nor denied here.

Nothing published, and the exposure is distinctive. The product's whole method is assembling operational data about a business from sources other than its own financial statements in order to estimate what that business actually earns. Nothing states what those sources are, what is retained, whether a scored business is told it has been scored, or what happens to the derived figures afterwards. A profile of a small firm's real turnover, built without that firm submitting it, is a sensitive artifact and no handling commitment attaches to it.

No security page, no certification of any kind, no trust portal, nothing on encryption, access control or testing. Complete absence rather than thin presentation. Recorded with the observation that this vendor sells to large European banking groups whose vendor security assessments would require far more than is published, so material almost certainly exists in private and none of it is public.

No licence, registration or supervisory relationship claimed, and the vendor sits outside the perimeter as a software supplier to licensed lenders. Worth noting for a later pass: a borrower facing lending platform is described as in development, and if the firm ever originates or holds credit itself rather than scoring for banks, both this axis and the standing operator test would need revisiting.

No fairness testing, no protected characteristic handling, no governance statement, and this is the most consequential C on the record. A model that estimates a small firm's real revenue from operational proxies rather than declared accounts is, by construction, making inferences about businesses whose formal reporting is incomplete, and the vendor publishes its own research on the shadow economy, which places the method squarely in markets where undeclared activity is common.

Proxy discrimination is the predictable failure mode: sector, location, informality and business type will correlate with owner characteristics, and nothing anywhere addresses it. A lender adopting this in the European Union would also need to answer the automated decision making questions for an applicant scored without submitting financials, and no material supports that.

Nothing published, and the recourse gap has an unusually concrete shape. A business declined here is declined on an estimate of its own revenue that it never supplied and cannot see. It has no way to know the figure the model produced, no route to correct the operational inputs behind it, and no stated appeal path, and the vendor's own framing is that no financial statements are needed, so the obvious remedy of submitting accounts is not part of the described process. Sharper than the standard lending case because the disputed fact is not the borrower's creditworthiness but the vendor's estimate of the borrower's basic finances.

A textbook instance of the standing rule that building permits silence: the models are described only as proprietary machine learning models, and nothing is named at any rung. No framework, no architecture, no feature families, no data providers, no cloud, no partner. Useful as the small vendor bookend to the same pattern seen on this roster's largest names, where Murex published an architecture, ION published the word algorithms and Linedata bought two AI companies and named nothing. The disclosure gap is not a function of vendor size.

Nothing published. No interface documentation, no core banking or loan origination system connectors named, no data ingestion description, no deployment architecture. That is a genuine gap rather than a formality for this product: a score derived from operational business metrics has to get those metrics from somewhere and has to return a decision into an origination workflow, and neither end is described. A buyer cannot tell what it would have to build.

Nothing published on either half. No hosting model, no cloud or on premise option, no residency statement, and no company address disclosed anywhere on the site, which is unusual in itself and relevant here: a bank in a regulated market buying a scoring service needs to know which jurisdiction processes its borrower data, and that cannot be determined from the vendor's own material.

No pricing, no pricing page, no commercial model stated. Every route is a demo request or a contact form. The one commercially shaped claim, more than 100 percent return per project, describes value rather than cost and is unattributed.

Eight named bank clients spanning large multinational groups and regional institutions, mostly across Central and Eastern Europe and Central Asia: Intesa, OTP Bank, Societe Generale, UniCredit, Jusan, Hamkor, Fast Bank and a Kyrgyz operator. More than 50 projects worldwide claimed.

Held at B rather than A because the breadth is in institution type rather than in what is sold to them: a single scoring product addressed at one segment, small and medium business lending, plus a consulting line. No enterprise, retail, capital markets or insurance coverage, and no evidence of the product being deployed across a group rather than in one subsidiary.