Fintech Systems · Pipeline Engineering · Behavioral Machine Learning
An autonomous operating system for consumer and commercial loan collections. Synthesizes Dagster asset-oriented pipelines, dbt transformations, LightGBM bounce scoring, and automated compliance guardrails enforcing statutory contact frequency and legal calling windows.
Loan accounts progress through discrete delinquency stages governed by an action-dependent transition matrix. The objective is to allocate scarce collector and digital channel capacity to minimize total portfolio loss:
Markov Decision Process for Debt Roll-Forward
At each billing cycle \(t\), portfolio accounts occupy states \(S_t \in \{C, \text{SMA}_0, \text{SMA}_1, \text{SMA}_2, B_1, B_2, B_3, \text{NPA}\}\). Dynamic allocation policy \(\pi(a \mid \mathbf{x})\) assigns treatment \(a \in \mathcal{A}\) (Voice Bot, SMS, Telecalling, Field Agent) to minimize total credit risk:
Subject to hard statutory constraints: legal calling windows \(H \in [08:00, 19:00]\), daily attempt caps \(K_{\text{day}} \le 2\), and absolute suppression for registered DNC or active Promise-to-Pay (PTP) accounts.
Test the dynamic allocation engine and regulatory compliance guardrails in real-time. Adjust borrower risk parameters, contact history, and time of day to observe automated policy evaluation and guardrail enforcement.
Evaluates borrower delinquency features against RBI & FDCPA statutory guardrails
Treatment Assignment
Rule: Bounced High Risk
Compliance Gate Status
All RBI / FDCPA contact rules satisfied.
Projected Roll-Forward Risk
Probability of migrating to deeper delinquency bucket next billing cycle