Finance & Banking/Team / Hackathon Prototype

Consent-First SME Underwriting Platform

A full-stack underwriting platform prototype that demonstrates a regulatory-inspired Account Aggregator workflow: moving from loan application to explicit SME consent, retrieval of synthetic bank transactions and GST records, deterministic financial ratio analysis, and an explainable credit cockpit for bank underwriters.

Prototype Disclaimer: Built as a hackathon prototype using synthetic data and simulated Account Aggregator responses. Does not connect to live banking networks, utilize trained machine learning credit models, or issue legally binding credit decisions.

Scoring ModelDeterministic 0–100Rule-based mathematical evaluation, not an opaque black-box LLM
Weighted Dimensions5 IndicatorsCash Flow (35%), Revenue (20%), Debt (25%), GST (10%), Anomaly (10%)
Profile Differentiation3 ProfilesBaseline (72/100), Healthy (94/100), Higher-Risk (41/100) on ₹8L request
Data IngestionSynthetic Bank + GST48–115 synthetic transactions & 12 months GST filing data per profile
01 / CONTEXT

Problem & Objective

The Problem

SME lending is bottlenecked by fragmented, opaque financial verification. Underwriters must reconcile disconnected bank statements, GST filing statuses, and debt burdens manually. The core challenge addressed is how authorized financial data can be ingested with explicit user consent and converted into an explainable preliminary credit assessment within a unified workflow.

Project Objective

Build a working end-to-end prototype featuring separate Bank and SME interfaces, backend-owned consent state management, synthetic data retrieval via a Mock Account Aggregator, and a deterministic 5-metric risk scoring engine with transparent explanations.

02 / OWNERSHIP

Role & Contribution

Team Contributor (Hackathon setting) — Worked across the full-stack architecture, contributing to the FastAPI backend routing, consent state synchronization, and the underwriter risk cockpit.

03 / METHODOLOGY

Analytical Approach

  1. 01

    Role-Separated Architecture: Developed a single Next.js frontend hosting distinct `/bank` (application creation & underwriter cockpit) and `/sme` (consent review & authorization) portals communicating with a unified FastAPI backend.

  2. 02

    Strict Consent Lifecycle: Built a backend-owned state machine (`PENDING` → `APPROVED` → `DATA_READY` / `REJECTED`) ensuring financial data is inaccessible until explicit SME consent is granted with time-limited (24-hour) expiry.

  3. 03

    Mock Account Aggregator: Implemented a mock AA service serving multi-month synthetic bank transaction histories (48 to 115 records) and 12-month GST filing datasets across three differentiated SME profiles.

  4. 04

    Deterministic Scoring Algorithm: Formulated an explainable scoring formula weighting five key financial dimensions into a normalized 0–100 score: Cash Flow Stability (35%), Revenue Stability (20%), Debt Health Proxy (25%), GST Consistency (10%), and Transaction Risk (10%).

  5. 05

    Explainable Underwriter Cockpit: Provided quantitative metric scores paired with plain-language risk flags (e.g. GST sales discrepancies > 20%, high debit ratios, anomalous keywords) and preliminary recommended credit limits.

04 / STACK

Verified Technologies

Frontend Architecture
Next.js 14React 18TypeScriptTailwind CSSLucide React
Backend & Services
PythonFastAPIUvicornPydanticRESTful API
Data & Risk Modeling
Synthetic JSON DatasetsIn-Memory State ManagementDeterministic Financial RatiosRule-Based Anomaly Flagging
Deployment & Version Control
VercelRenderGitGitHub
05 / INSIGHTS

Verified Findings

Predictable Profile Stratification Under Rule-Based Engine

For an identical ₹8,00,000 synthetic loan request, the deterministic engine produced distinct assessments across test datasets: SME001 (Baseline) scored 72/100 (₹5,76,000 limit); SME002 (Healthy) scored 94/100 (₹7,52,000 limit); SME003 (Higher-Risk) scored 41/100 (₹3,28,000 limit) due to irregular cash flows and low debit coverage.

Multi-Source Reconciliation Identifies Concealed Discrepancies

The GST consistency module cross-verified monthly bank credits flagged as 'SALES' against reported GST sales. Months exhibiting discrepancies > 20% triggered automatic audit flags, demonstrating how dual-source verification catches reporting mismatches.

Consent-Gated Data Privacy by Architecture

Enforcing consent state strictly at the backend prevented premature data leakage. Underwriter endpoints returned 403 Forbidden on data requests until the SME explicitly authorized access via the portal.

07 / REFLECTION

What I Learned

  • Designing financial underwriting workflows where data privacy and explicit user consent are first-class architectural constraints.
  • Translating qualitative credit risk factors into deterministic, explainable mathematical formulas.
  • Cross-verifying transactional banking data against governmental filing records (GST) to flag reporting variances.
  • Collaborating across frontend, backend, and risk-engine modules under a shared API contract in a fast-paced hackathon setting.