Decision Accountability
Infrastructure for
Indian Lending.
NomoCrit is the model risk management layer for RBI-regulated NBFCs. Every AI credit decision becomes an immutable, explainable, regulator-ready record — built for the FREE-AI framework and the RBI’s 2026 draft MRM guidance.
FREE-AI (RBI) Principle-Aligned DPDP-Aligned by Design Audit-Trail Native Data Resident in India VPC-Isolated
- [ RBI/2022-23/111 ]
- [ DPDP ACT 2023 ]
- [ FREE-AI FRAMEWORK ]
- [ RBI DRAFT MRM GUIDANCE 2026 ]
- [ FAIR PRACTICES CODE ]
India’s lending AI is making consequential decisions without accountability infrastructure.
Between 2022 and 2024, Indian NBFCs and digital lenders deployed ML-based credit scoring at a pace that outran the governance infrastructure available to support it. The RBI Digital Lending Guidelines required lenders to communicate the grounds of rejection to applicants and to maintain model risk management frameworks adequate to the complexity of the models in use. What lenders actually had, in most cases, was a score, a binary decision, and no documented pathway between the two.
Millions of credit decisions are being made each year by models that are, from a regulatory standpoint, functionally opaque — not because the technology for explainability does not exist, but because no one has built the layer of infrastructure that would make explainability native to the decisioning workflow rather than a retrospective exercise conducted under audit pressure.
Models drift silently across credit cycles
No mechanism alerts lenders when post-approval model performance diverges from baseline as the rate cycle, bureau format, or monsoon cohort shifts.
Grounds-of-Rejection notices lack legal standing
Rejection communications rarely satisfy the specificity required under the RBI Digital Lending Guidelines (RBI/2022-23/111) and the Fair Practices Code.
Fairness metrics are reported, not enforced
Lenders surface demographic-parity scores in audits but have no automated recalibration when thresholds are breached.
Not a model. Not a tool.
A decision governance substrate.
Infrastructure doesn’t ask to be evaluated on features. It asks to be evaluated on position. NomoCrit’s position is between the raw output of any credit-scoring model and the regulatory surface — the point at which a probabilistic score becomes a consequential decision.
A product is bought, deployed, and replaced. Infrastructure is the layer everything else depends on. NomoCrit does not compete with your scoring model, your LOS, or your CRM — it governs the decisions those systems produce. When models are retrained, NomoCrit continues. When regulation changes, NomoCrit adapts. When an RBI auditor arrives, NomoCrit generates the artifact.
The question it answers is not “how good is your AI?” It is “can you prove your AI is accountable?” These are different questions. Only one has a yes-or-no answer that survives regulatory scrutiny.
Explainability Engine
Every credit decision is decomposed into a per-feature SHAP contribution map, surfaced as both a machine-readable JSON artifact and a human-readable grounds-of-rejection notice.
Adaptive Threshold System
Approval thresholds recalibrate automatically when population-level drift is detected across a rolling credit window — without requiring manual model retraining.
Fairness Monitoring
Demographic-parity and equalised-odds metrics are computed on every decision batch, with automated alerts when any protected-class disparity exceeds defined tolerance bands.
Recourse Pathway Infrastructure
A structured appeal flow accompanies every grounds-of-rejection artifact — providing applicants with a compliant recourse pathway and lenders with a defensible audit record.
From raw decision to auditable governance
— in one pass.
Stateless by default
Each decision evaluation is a pure function of its inputs. NomoCrit holds no mutable model state — eliminating a class of silent drift that plagues stateful scoring pipelines.
Every decision is a ledger entry
Decisions, explanations, threshold states, and fairness metrics are written to an append-only log at evaluation time — not reconstructed on demand for auditors.
Threshold drift triggers recalibration
When population-level statistics shift beyond tolerance, the threshold recalibration pipeline activates automatically — producing a new threshold set with a full provenance record.
Built for the regulatory reality
of Indian AI in credit.
An applicant is rejected. What happens next?
NomoCrit intercepts the decision signal before it reaches the applicant and generates a structured grounds-of-rejection artifact — SHAP-grounded, mapped to the Fair Practices Code categories, with a documented recourse pathway attached to every record.
RBI asks for a model audit. What can you produce?
A decision ledger that already exists as a primary operational output — every decision linked to its model version, threshold log, SHAP vector, and fairness metrics, exportable as a structured submission without manual reconstruction.
Your model drifts post-monsoon. Who catches it?
NomoCrit's adaptive threshold engine monitors four drift signals in parallel — PSI, feature-level distribution shift, fairness deviation, and segment approval-rate change — and triggers recalibration before drifted decisions accumulate into a documented discriminatory pattern.
Accountability is not a value; it is a latency specification.PRINCIPAL NOTE — NOMOCRIT ARCHITECTURE BRIEF
Integrates with the Indian lending stack
— not against it.
Additive, not disruptive
NomoCrit wraps existing scoring pipelines. There is no model migration, no re-integration of bureau APIs, and no LOS reconfiguration. It inserts at the decision point.
Output format agnostic
NomoCrit consumes any score on a 0–1 probability scale. It emits standard JSON artifacts and webhook events — consumable by any downstream LMS or reporting system.
RBI reporting included
Audit exports are formatted to satisfy RBI inspection requirements out of the box. No post-processing, no manual annotation, no last-minute restructuring for an inquiry.
Built at the intersection of machine learning, financial regulation, and the scale of Indian credit.
India’s digital lending infrastructure grew faster than its governance infrastructure could follow. By 2024, AI was making millions of consequential credit decisions per day with no standardised explainability requirement, no enforceable fairness baseline, and no audit mechanism that a regulator could independently verify. NomoCrit was built specifically for this moment.
We believe responsible AI is not a feature that can be layered on after deployment. It is infrastructure — the kind that must be present at the decision point, operating at decision latency, with a complete record of every evaluation. A model that cannot explain itself is not a governed model. A lender that cannot produce a decision audit trail does not have a governance programme — it has a compliance deck.
The version of Indian lending we are working towards is one where every AI decision is explainable, every recourse pathway is open, and every audit is producible in hours — not because of regulatory pressure, but because the infrastructure makes accountability the default. That is the only version of AI in credit that scales.
NomoCrit began as credit-risk governance research at IIT Delhi, where Advait co-developed DTPG, RVDR, and CFDI — decision-governance metrics for evaluating how lending models behave under drift and disparity. NomoCrit is that research, hardened into production infrastructure for the Indian lending stack.
“Accountability is not a competitive advantage. It is the baseline.”Get in touch.
If you’re building accountable lending infrastructure, we want to hear from you.