Decide in real time, with the evidence attached.
Nyx is a decisioning engine for banks, NBFCs, fintechs and PSPs. It approves, reviews or declines inside the authorization path, whether the buyer is a person or an agent acting on their mandate, and every verdict carries the evidence to defend it.
// response · 87ms
{
"action": "review",
"score": 0.91,
"severity_band": "high",
"reason": "velocity + geo mismatch",
"evaluators": [
{"rule": "cnp_velocity", "hit": true},
{"model": "txnCBM@v14", "score": 0.91}
],
"audit_ref": "dec_EXAMPLE…",
"latency_ms": 87
}
Two ends of the same transaction.
The engine is one. The problem it solves depends on which end you sit at.
Every decision you have to defend.
- Approve, review or decline inside the authorization path Pilot now
- Every verdict carries its reason, rule hits and audit reference Pilot now
- Verify an agent's identity, principal, mandate, scope and caps Built with you
- Regulatory evidence packs generated by the engine Built with you
- Credit risk, underwriting, disputes and AML Built with you
Agent-driven buyers, without the declines.
- Tell a verified agent from unverified automation and from a person Built with you
- Check the mandate the agent presents: scope, caps, expiry Built with you
- An evidence record for every agent purchase, for the dispute that follows Built with you
- Publish products, availability and policy to the agent protocols Pilot now
Not on the list? Bring us the decision you actually need to make.
Three forces are compressing the decision window at once.
Attacks are produced faster than models retrain, regulators are putting real-time controls on the party making the decision, and a new kind of customer has started transacting on its own with nobody yet liable for it. Fraud is where most decision stacks break first, but none of these three stops at fraud.
The attack is automated.
Deepfake KYC, synthetic identity and GenAI social engineering, generated at a scale supervised models cannot retrain against.
The liability is moving to you.
Regulators are putting real-time controls, reporting and reimbursement on the party making the decision, on dated timelines. The nearest one is a risk-based authentication requirement, not a recommendation.
The customer may not be human.
Agents find, evaluate and buy on a person's behalf. The device and behavioral signals every stack leans on vanish with them, unverified agent traffic falls back to card-not-present treatment, and no rail has yet said who is liable.
Force 2, by regime · IN · RBI AUTHENTICATION DIRECTIONS 2025, CROSS-BORDER CNP RISK-BASED AUTHENTICATION FROM 1 OCT 2026 · DPIP LIVE 1 SEP 2026, PHASE 2 SCORES BEFORE COMPLETION · RBI OUTSOURCING DIRECTIONS NOV 2025 · DPDP OBLIGATIONS FROM 13 MAY 2027 · AE · CBUAE 3057, FINES SINCE 31 MAR 2026 · CBUAE AI GUIDANCE 11 FEB 2026 · SA · SAMA CFF IN FORCE 13 APR 2026 · EU · VERIFICATION OF PAYEE (IPR), PSD3/PSR LIABILITY SHIFT 2028 · UK · APP REIMBURSEMENT IN FORCE 7 OCT 2024 · SG · MAS SHARED RESPONSIBILITY FRAMEWORK · AU · SCAMS PREVENTION ACT 2025 · US · NO FEDERAL MANDATE YET, LIABILITY PRESSURE BUILDING
One engine. Six decisions. An honest status on each.
Nothing domain-specific is welded into the kernel. A versioned pack declares the entities, features, policies and actions a domain needs, so the second decision is a configuration exercise rather than a second procurement.
Five of these are decisions a bank, NBFC, fintech, processor or lender makes about a customer. Agentic commerce is the one a PSP runs for its merchants, on the agents now buying on a customer's behalf.
Approve, review or decline a payment or account action, inside the authorization path.
OPEN TO DESIGN PARTNERSDecide on the agent that buys: its signature, its principal, its mandate. Publishing to the agent protocols is ready to pilot today.
OPEN TO DESIGN PARTNERSExtend, limit, price or decline exposure, with reasons a customer can be given.
OPEN TO DESIGN PARTNERSAccept, refer or decline against written policy, with the reasoning attached.
PARTNER-LEDChargeback and claim outcomes, where assembling the evidence is most of the work.
PARTNER-LEDScreening, onboarding risk and transaction monitoring, audited end to end.
Two are ready to pilot today: fraud decisioning for institutions, and publishing to the agent protocols for the merchants a PSP serves. Any of these can be the one we start with. Compare all six →
The verdict stays fast. The investigation runs beside it.
One synchronous call returns the action, score, reason, rule hits, model version and audit reference, and writes the evidence record beside it. The heavier AI work runs asynchronously and never enters the authorization path.
† Design target to validate in the first pilot, not yet independently benchmarked. Architecture, primitives and the decision API →
How it deploys is your call, not our constraint.
The same engine, the same API and the same evidence record in all three models. Pick the one your regulator, your security team and your stage of growth allow.
We run it in the region you pick, single-tenant by default. Multi-tenant where one institution serves its own merchants on it. The fastest route for fintechs, PSPs, merchants and smaller institutions with no platform team to spare.
Deployed into your own AWS, Azure or GCP account and VPC. Your keys, your network, your data residency; we ship and support the release.
On-premises inside your perimeter, for institutions whose regulator or policy requires it. No bulk extraction to a third-party black box.
Your data stays yours, in-country, in every model. Native data-lake integration, PII masking, injection guardrails on every AI call, and models trained on your transactions only.
Role-based access and enterprise SSO (Keycloak / OIDC). An evidence record per decision, idempotent and replayable, structured for regulatory reporting.
SAMA CFF and CBUAE 3057 control mappings are written. The evidence pack for CBUAE's AI guidance (11 Feb 2026) and FREE-AI is shaped with the first partner under each; other regimes are mapped with the first partner in that jurisdiction.
Open standards end to end. Your models, features and decision history stay portable, and exit terms are written into the pilot, not discovered later.
Full architecture, governance and operating posture on the Platform page →
A team that has built the systems businesses decide with.
NaxVerse is a senior engineering team spanning real-time authorization, fraud scoring, risk decisioning, ML infrastructure and DevSecOps. People who have built and operated these systems inside banks and payment processors, now building the platform they wished they had.
Led by founder Umesh Kolhe: two decades across the payments lifecycle at JPMorgan, TSYS and Pine Labs, from authorization and fraud scoring to the ML infrastructure behind them.
NaxVerse Technologies Pvt Ltd is headquartered in Noida, India, in the Delhi NCR, and works with banks, NBFCs, fintechs, processors, lenders and merchants wherever they are regulated. Founder-led, pre-launch, and forming a small design-partner cohort now.
Press & due-diligence pack →Advisory group forming across fintech, payments and risk
The questions that come up first.
Short answers, carrying the same qualifiers as the rest of the site.
What is Nyx?
An AI-native decisioning engine. It approves, reviews or declines a transaction or an account action in real time, inside the authorization path, and attaches the evidence to defend that decision afterwards. One engine covers fraud, agent trust, credit risk, underwriting, disputes, AML and KYC through versioned domain packs, instead of one vendor per decision.
What is the evidence that gets attached to a decision?
Every decision returns the action, the score, the severity band, the policies and rules that fired, the model version and an audit reference in the same response. A canonical evidence record is written beside it: actor, principal, mandate, what was shown, consent, decision, outcome and dispute status. That record is what a dispute or an examination is answered with later. The record is in build, and its schema is versioned from the first write.
Who is it for?
Banks, NBFCs, digital banks, fintechs, payment aggregators and PSPs, acquirers, lenders and merchants. Institutions are the direct buyer. Merchants are reached through the PSPs that embed the trust endpoint for them, and a merchant can also come directly. Size is not a gate, because the hosted deployment removes the need for an in-house platform team.
Is Nyx running in production today?
No. NaxVerse is pre-launch and founder-led. The fraud pack is built and ready to pilot, agent trust is open to design partners, and the performance figures on this site are design targets to validate in a first pilot rather than benchmarked results. NaxVerse publishes targets, not testimonials it does not have.
Where does the data live?
In your jurisdiction, in all three deployment models: hosted and managed by NaxVerse in the region you choose, inside your own cloud account and VPC, or in your data center. The engine, the API, the packs and the evidence record are identical in all three. Models train on your data only, and no deployment requires bulk extraction.
How does an engagement start?
With a shadow replay of your own decision history, run in the deployment model you choose with nothing leaving your perimeter. It produces four numbers you do not have today: false declines recoverable, mule accounts missed, agent-originated traffic share, and dispute cases defensible with evidence.