HyperVerge

The removal test leaves nothing, and the company's origin explains why. It began as a computer vision team that found success in academic competition before turning the work on identity, and the chief executive's account of the founding problem is that human verification of face images was too slow to scale.

Every function is a model: face matching, passive and active liveness, deepfake detection, optical character recognition across a long list of government document types, forgery and tampering detection, and photograph based verification of a business premises from an image of its signage and inventory. There is no marketplace, orchestration engine or rules layer underneath that would survive, which is what distinguishes this from Signzy and Alloy in the same lane.

One regulated human step is supported and the rest is automated. The video customer identification process the platform implements requires a live interaction between the applicant and an official of the institution, so building for that modality means the product accommodates a human decision maker where the rule demands one.

Everything else runs without a person: identity authority based electronic verification, document capture and liveness complete in under a minute by design, and the stated benefit throughout is removing manual review. No review queue, escalation threshold, alternative evidence route or manual adjudication surface is described for an applicant the automated path cannot verify, which is the specific gap against Persona's step up mechanism.

More measurement exists here than almost anywhere in this lane. An accuracy figure of 99.5 percent is published, which most competitors decline to state at all, and the core face model has been independently benchmarked against a government standard and separately evaluated in a competitive national procurement, so a third party with no commercial interest has tested it. That combination is what earns the grade.

The qualifications are real: the accuracy figure carries no methodology, test set or breakdown by document type or population, the accompanying claim to be highest among competitors is unsubstantiated comparison, and the benchmark results themselves are referenced rather than published. Nothing at all is measured for the newer lending layer, where financial spreading and forgery detection feed credit decisions.

Volume, named customers and an unusual financial fact all support this. More than a billion identities verified across over 200 customers, with roughly 30 million customers onboarded monthly, which the company frames against a national population as approximately 3 percent of India each month.

Named institutions include the largest public sector bank, a major non bank lender and insurer, a large diversified financial group, a Malaysian bank and an international consumer lender, with executives quoted on reduced cybercrime complaints and chargebacks after go live. Published revenue reached 149 crore rupees in the most recent financial year.

The standout is capital efficiency: roughly 19 million dollars of annual recurring revenue built on about 1.1 million dollars raised, which means the business was validated by customers paying for it rather than by investors underwriting it, and that is a different and stronger kind of evidence than a funding round.

No data boundary statement was located. The models are the company's own and the corpus behind them is implied by scale rather than described, since a billion verifications across 200 customers is exactly the kind of accumulation that improves computer vision, and nothing states whether images and documents submitted by one institution's applicants train models serving another.

That question is sharper here than for a rules based product because model quality in this category is a direct function of training breadth, so the commercial incentive to pool is strong and the disclosure is absent. Compare Rulebase, which forecloses training on customer data in a single sentence.

No data protection agreement, retention schedule, subprocessor list or deletion commitment was located, and the data types involved are unusually broad even for this lane. Biometric captures run at roughly 30 million onboardings a month, and the lending layer adds screening against court records and police first information report databases, which means criminal justice records about named loan applicants are processed alongside identity documents, income evidence and credit bureau data.

Nothing published addresses how long a facial template persists, how a police record match is stored, or how the home market's data protection statute applies across a footprint spanning many jurisdictions.

End to end encryption is asserted and no attestation, certification, trust centre or enumerated framework was located. The government benchmarking the company has undergone tests model performance rather than organisational security controls, so it does not substitute, and no accredited presentation attack detection certification appears, which is the specific assurance a buyer compares in this lane and which two peers publish. For a platform holding biometric captures at this volume the absence of any published assurance set is the clearest gap in an otherwise unusually well evidenced profile.

Grounded in named regulated processes and a competitively awarded government mandate, which is the pattern now recorded across every Indian vendor in this index. The platform implements the central bank's video customer identification process, a specifically authorised onboarding modality with prescribed requirements, alongside identity authority based electronic and offline verification, and it addresses the securities regulator's separate know your customer guidelines for brokerage clients.

Beyond compliance, the company was selected under the national artificial intelligence mission as one of only two winners from more than twenty five competing teams for a national public examination face recognition contract, which is a government evaluation passed rather than a standard met. Sixth A on this axis after OnFinance AI, Akur8, NICE Actimize, Perfios and Signzy.

The second B on this axis in the identity lane, and it is earned on the same reasoning recorded for Incode. The company's face recognition has been independently benchmarked against United States homeland security standards, and it was selected under a national artificial intelligence mission after competitive evaluation against more than twenty five teams for a face recognition contract.

Submitting a face model to a government evaluation is a genuine adversarial test, because those benchmarks measure demographic differentials by design, and a vendor that enters one has accepted external measurement of exactly the error profile this axis exists to expose. What keeps it off an A is the same limit as Incode: the demographic breakdown those submissions generate is not published.

Two newer exposures also sit outside the face benchmark entirely, since court and police record screening across name and address variations carries a matching error profile of its own, and a multiple identity flag is an accusation against a named applicant.

No guarantee, indemnity or falsifiable commitment binds the vendor to a verification outcome. The stated mission is enabling people excluded from finance to reach it, and the product exists to reduce the friction that causes applicants to drop out, so the commercial incentive is genuinely aligned against wrongful rejection. That is alignment rather than recourse.

An applicant whose face fails to match, whose document is judged tampered, who is flagged as appearing under multiple identities or who matches a police record is not told which system produced the finding, cannot see the evidence, and has no described route to correct it, and the last two are adverse findings that follow a person well beyond the application they were made in.

The model half of the chain is short by construction and the company says so, since the computer vision stack was built in house from its founders' own research rather than assembled from third party engines, which is the property that earned Incode its position on this axis. The data half is partly named: the national identity authority, court and police record databases and credit bureau reports are all identified as sources feeding verification and lending checks. What is not disclosed is the specific providers behind those categories, no sanctions or watchlist data supplier is named, and no subprocessor list or hosting arrangement appears anywhere.

The integration surface is broad and specific on the input side. Application programming interfaces sit alongside mobile and web software development kits and a no code deployment path the company positions as removing heavy information technology dependency, with go live stated in hours.

Supported document types are enumerated rather than implied, covering the national identity number, tax identifier, passport, driving licence, voter identity, business tax registration and vehicle registration, and the identity authority integration is direct.

What is not evidenced is the downstream side: no core banking system, loan origination platform or customer relationship system is named as an integration target, so a bank cannot establish how a verification result and its audit trail land where its own staff work.

No hosting provider, region selection, residency commitment or private deployment option was located. Coverage claimed across more than 200 countries with operations spanning India, the United States and Southeast Asia means biometric captures and identity documents cross many transfer regimes, and the home market applies its own localisation rules to financial data. Nothing published tells an institution where a facial template or a document image processed on its behalf comes to rest, or whether that location can be constrained.

No pricing, packaging or basis of charge is published and every route in is a demo request. The company describes a consultative approach in which it works on a client's business problem rather than selling an application programming interface or software development kit off the shelf, which is a sales posture rather than a pricing disclosure and implies negotiated commercial terms. Nothing indicates whether charging is per verification, per active customer or as a platform subscription.

Within financial services the coverage is genuine and separately maintained, with distinct propositions for banks and non bank lenders, for securities brokerages under their own regulator's rules, and for insurers, plus a lending specific product layer covering underwriting document work rather than onboarding alone. Geographic reach is stated across more than 200 countries with offices in India, the United States and Southeast Asia and growth in Africa.

What holds this at B rather than higher is that financial services sits alongside telecommunications, gaming, electronic commerce, logistics and education technology as one vertical among several, so the platform is a general identity product with a strong financial services practice rather than a purpose built financial institution product.