FIVIby Xivi Innovations

For banks and NBFCs

Check every AI credit score before you act on it.

FIVI runs next to your existing credit, AML or risk model. For each live decision it returns Cleared, Abstain or Flagged, with a reason code and an audit record.

Request a pilot See how it works

Founded by Chinmaya Naik and Abhinav Navneet · Team

Shadow modeRead-only
Your model→FIVI→Decision
ApplicationTwo-wheeler loan · #A-20418
Model score0.71
FIVI verdict
AbstainCovariate drift

Income-linked inputs have shifted past their calibrated limit.

Drift vs. limitElevated
Action
Routed to a credit officer
Audit
SHA-256 · 3bfb01d7…

Illustrative example. Policy, thresholds and error limits are set by each lender.

Tested offline on
LendingClubNSSO AIDISUCI CreditEllipticBinance UST
Aligned with
RBI FREE-AIDPDP Act 2023RBI IT Outsourcing (NBFCs)

Problem

Credit models drift, and nothing tells you.

A model learns from past data. When borrower cash flows or the economy change, its real error rate rises while its scores still look confident.

17.4%→23.3%

Default rate on LendingClub loans, 2013–14 versus 2016. Defaults rose 34% under the same approval cut-off, and the model gave no warning.

9.9 crore

Digital personal loans sanctioned in India between April and December 2025.

34+ crore

Account Aggregator consents fulfilled, which means more cash-flow data and faster regional and seasonal shifts.

Sources: FIVI LendingClub benchmark on resolved historical loans; The Economic Times; Sahamati (Dec 2025).

Existing tools

None of them check each decision.

Scoring engines

Produce a score. They do not check whether their error limits still hold.

Model monitoring

Alert a team after the fact while applications keep flowing.

AML alerting

Sends every alert to a manual review queue.

Model governance

Reviews models periodically, not live decisions.

FIVI adds a check on each decision. It uses only the scores and inputs you already have and needs no access to the model's internals.

Solution

Three possible verdicts.

For each decision, FIVI checks whether the score can be trusted today, within your stated error limit.

Cleared

The score is within its error limit. Automate the decision.

Abstain

Uncertainty is high or inputs have drifted. Send to a credit officer.

Flagged

A hard limit is breached. Escalate for review.

Reason codes: Band uncertaintyCovariate driftSparse subgroupThreshold conflict

A breach stops automation until a risk officer resets it.

The circuit breaker watches the worst-performing subgroup, so one drifting segment stops automation even when the overall average looks fine. Only a named risk officer can restart it.

Technology

A proprietary core, designed to resist reverse engineering.

The calibration at the heart of FIVI is kept private. It cannot be copied from published research.

Provisional patent filedCovers the system architecture.
Trade-secret calibrationParameters are never disclosed.
Secure enclave in designVerification without exposing internals.

Integration

How FIVI connects.

FIVI runs in shadow mode inside your own cloud. It reads a copy of your model's output and inputs, checks them in parallel, and never writes back to your lending system.

What FIVI reads

  • Your model's score or output
  • The input features the model already uses
  • A historical calibration sample and a small set of audit labels

What FIVI does not need

  • Model weights, source code or retraining
  • Demographic or caste data at origination
  • Stored customer PII
  • Write access to lending ledgers

RBI FREE-AI

Uncertain cases go to a person, every verdict has a reason code, and every decision is logged.

Data minimisation

No new borrower data is needed, in line with section 6(1) of the DPDP Act 2023.

Data in India

FIVI runs in the lender's domestic cloud, consistent with the RBI IT outsourcing directions for NBFCs.

Results

Offline tests on five datasets.

These are retrospective benchmarks on public and survey data. They are not live lender deployments.

LendingClubUS consumer credit · 2013–2016
+34%default spike absorbed

Error limits held while defaults rose. Uncertain cases went to officers.

NSSO AIDISIndia household survey · 170,000 households
15.1% → 8.8%group exclusion gap

Standard grouping missed one group. A small offline audit closed the gap without using sensitive attributes at origination.

UCI CreditTaiwan · 30,000 card accounts
9.8% = 9.8%missed defaults, both sexes

Missed-default rates matched across sexes with sex withheld from the model.

EllipticOn-chain AML · 203,769 Bitcoin transactions
0.48 → 0.84detection AUC

AML detection recovered after a market shift.

Binance UST2022 stablecoin collapse · price data
44 hrsearly warning

Warned before the peg broke, with no false alarms on the BUSD benchmark.

Sources: LendingClub; NSSO AIDIS; Yeh & Lien (2009), UCI; Weber et al. (2019), Elliptic; Binance market data.

Pilot

A 12-week shadow pilot on your own data.

We are starting with mid and upper-layer NBFCs that run automated credit models, in two-wheeler, farm-equipment, rural and MSME lending. Your live decisions are not touched.

Weeks 1–4Deploy and scopeCloud setup, data schema, read-only access and risk tolerances.
Weeks 5–8CalibrateBaseline performance, error margins and drift thresholds on past data.
Weeks 9–10Shadow runRun alongside live traffic and measure verdicts and referral rates.
Weeks 11–12ReviewA model-risk validation dossier, audit logs and a rollout plan for your risk committee.

Team

Who we are.

CN

Chinmaya Naik

Co-Founder and CEO

Computer Science engineer with 7+ years in product development and management.

AN

Abhinav Navneet

Co-Founder and CTO

Computer Science engineer with 5+ years in technology, research and IP.

HT

Harsh Trivedi

Adviser, Finance

ACCA graduate and Director at Prepper Gurukul, with 8+ years of experience.

Contact

Talk to us.

Tell us about your credit, AML or risk workflow and we will scope a read-only pilot.

xivi.tech@gmail.com