GiniMachine

Model building is the entire product rather than a feature of one. The platform ingests a lender's historical loan performance, constructs predictive scoring models automatically using decision tree methods, validates and deploys them in seconds to minutes, and monitors their behaviour afterwards. Remove the machine learning and nothing remains, since the company sells neither a workflow system nor a data source but the model itself.

Control over the consequential setting stays with the lender, which tunes its own cut-off value and defines the risk level it will accept, and the platform supports rule-based decisioning alongside the model so policy can override score. Against that, the company describes the system as absolutely automated and autonomous, not requiring constant human assistance, and promises decisions in seconds. Held at B because those two positions are not reconciled anywhere, and no referral, review or exception path is described for applications near the cut-off.

More method is disclosed than most peers offer. The technique is named as decision tree based with automated construction, which matters because tree models are inherently more inspectable than neural approaches, validation is described as part of the build sequence rather than an afterthought, and the platform monitors its own deployed models. The company positions itself directly against black-box tools.

Held at B because no accuracy measure, discrimination statistic, validation methodology or monitoring threshold is published, so transparency is claimed structurally rather than evidenced numerically.

One institution is named, a non-bank financial company in Mongolia whose chief business officer is quoted, alongside a second customer application performing small business assessment and a product lead quoted on time saved of roughly twenty minutes per application against a prior process taking well over thirty. Portfolio claims of higher acceptance rates and materially reduced non-performing loans appear without figures, baselines or attribution. The record is thin for a platform whose output is credit decisions.

One structural feature helps: models are built from each lender's own historical data and belong to that lender, so there is no shared pooled model across the customer base. Held at C because no boundary statement accompanies it and nothing states whether the automatic model construction method improves from patterns observed across customers, or what happens to a lender's data and models after an engagement ends.

No data protection agreement, retention schedule or subprocessor list was located. The platform scores applicants using rental and utility payment records, asset ownership and public records alongside bureau data, which is a wide personal data footprint assembled about people who may not know those sources were consulted, and none of the handling terms are published.

No attestation, certification, trust centre or enumerated control set was located. Lenders upload complete historical loan books including borrower attributes and repayment outcomes in order to train models, which is among the more sensitive datasets an institution holds, and no control documentation accompanies the upload.

No regulator, statute, supervisory expectation or lending rule is named anywhere in published material, in any of the several markets the platform serves. For a tool that builds and deploys credit scoring models, the absence is notable, since model documentation and adverse action requirements differ across those jurisdictions and a lender adopting a no-code model builder inherits all of them.

The company claims its platform eliminates human bias in evaluating applicants, and that claim does not survive examination of how the product works. Models are trained on the lender's own historical lending decisions, so any pattern in who was previously approved or declined is learned and then applied consistently at speed.

An independent comparison of this category states the problem precisely: train a model on biased history and it repeats that bias, so what makes AI scoring defensible is the fairness testing, monitoring and reason codes wrapped around it rather than the algorithm itself. No fairness testing, disparity monitoring or reason code capability is described.

No guarantee, indemnity or correction process was located. The declined applicant is unaddressed, and the gap is compounded by the no-code positioning: a lender with no data science team may be unable to explain why its own model declined someone, since the model was constructed automatically and no reason code capability is described.

The modelling technique is the company's own and is named at method level, which removes any external model dependency, and the automatic construction logic is described only as a proprietary element. External data sources are characterised as multiple credit bureaus and alternative sources including rental, utility and public records without any provider being identified, so the inputs determining who scores well remain undisclosed.

The platform is integrated into its parent company's lending suite, which gives it a route into origination workflow rather than sitting apart from it, offers interface access for connection to other systems, and is described as integrating with multiple credit bureaus. Held at B because not one bureau, origination system or core platform is named individually, so a buyer cannot confirm whether its own data sources are supported.

Delivery is described as a cloud-hosted web application with no hosting provider, region selection, residency commitment or private deployment option stated. That matters more than usual given customers named or implied across five widely separated jurisdictions, several with data localisation requirements covering credit data.

A free trial is offered, which lets a buyer reach the product without a sales process, and no pricing, packaging or basis of charge is published. For a platform targeting smaller lenders explicitly because they lack analysts, cost relative to hiring one is the comparison that decides the purchase and it is unsupported.

Lending coverage is broad, spanning online, commercial, point-of-sale, auto and credit card lending, plus small business, merchant cash advance, trade finance, factoring, leasing and working capital finance through its parent's lending suite, with regional presence indicated across the United Kingdom, Saudi Arabia, Australia, Canada and the Philippines. Buyers include banks, non-bank financial institutions and non-financial businesses entering lending for the first time. Held at B because presence in those markets is asserted through the parent rather than evidenced directly.