Steady Rabbit
Get in touch

AI-Powered Underwriting Platform Transformation

Cutting Loan-Approval Time by 50 % and Lifting Risk-Scoring Precision 12 pp in 24 Weeks with Steady Rabbit’s Core-Flex Micro-GCC Model

Client
Confidential
FinTech start-up
Mission

Automate SME loan underwriting with AI-driven document intelligence & risk scoring

Pre-Engagement State

Python ETL scripts, Flask API, PostgreSQL, manual PDF uploads

Confidential FinTech
start-up


Case Study

Executive Summary

A venture-backed US–India FinTech scale-up set an audacious target: shrink commercial-loan underwriting from five days to < 48 hours—without compromising regulatory rigor. Their prototype (Python + Excel) proved market demand but collapsed under messy documents, model-ops drift, and looming SOC 2 scrutiny.

Steady Rabbit mobilised a six-person Core-Flex Micro-GCC squad (plus Buffer insurance) and, in 24 weeks, delivered an AI-ready underwriting platform that:

  • Ingests 2 TB/day of multi-format documents via an event-driven pipeline
  • Uses GenAI embeddings to cut manual review 50 %
  • Lifts model AUC 0.72 → 0.84 (+12 pp risk precision)
  • Passes SOC 2 Type II & US/EU lending audits three months early
  • Auto-scales for 5 × booking spikes while trimming GPU spend 35 %
  • Unlocks US $ 7.8 M ARR within 30 days of launch

Every sprint landed on time, helping the company secure an US $ 18 M Series B at a premium valuation.

Client Profile & Business Context

  • Client
    Confidential FinTech start-up

    (HQ Delaware; Indian engineering hub)

  • Founded

    2019

  • Mission

    Automate SME loan underwriting with AI-driven document intelligence & risk scoring

  • Pre-Engagement Stack

    Python ETL scripts, Flask API, PostgreSQL, manual PDF uploads

  • Market Focus

    Community & regional banks (US Midwest) and Indian NBFCs

Scale-Up Imperatives

Document Intelligence

Messy tax returns, bank PDFs, scanned invoices

Model-Ops Discipline

Versioning, retraining, regulator explainability

SOC 2 Compliance

Non-negotiable for US
lenders

Predictable Go-Live

Conditional LOIs worth US $ 7.8 M hinged on a six-month launch

Problem Statement / Key Challenges

Challenge

Unstructured document chaos (OCR 72 % accuracy)

Impact if Unsolved

Manual QA doomed SLA; target < 48 h

Challenge

No model pipeline

Impact if Unsolved

Feature tweaks took weeks; risk of drift

Challenge

SOC 2 & FDIC audit looming

Impact if Unsolved

Contract kill-switch if controls failed

Challenge

Elastic demand spikes (5 × quarterly)

Impact if Unsolved

PoC infra costs tripled under load

Challenge

Five generalists in-house

Impact if Unsolved

Couldn’t cover security, MLOps, DevOps simultaneously

Steady Rabbit’s Approach

Micro-GCC Squad

Layer
Roles & Size
Mandate
Core (6 FTE)
Squad Lead/PO, 2 Data/ML eng, Go micro-services eng, DevOps/SRE, QA auto
Own backlog & ship end-to-end platform
Flex (2 SME)
CISSP cloud-security architect, GenAI engineer (LangChain/embeddings)
High-risk spikes: SOC 2 controls, embeddings
Buffer (1)
Shadow data engineer (vendor-funded)
PTO/attrition cover in < 4 h

Shift-Left Governance

  • 7 Plan-Left gates baked into Jira (Persona → Test Note).
  • Card can’t enter Dev-Doing until gates green (avg < 28 min).
  • SteadCAST tracks Risk-High WIP %, test-note coverage, capacity vs velocity.

Discovery Sprint 0 – Outputs

  • STRIDE threat model + compliance matrix (SOC 2, FDIC, GDPR)
  • North-Star architecture: Kafka ingest → OCR/embeddings → Feast feature store → Go risk-service
  • Baseline KPIs: SLA < 48 h, AUC ≥ 0.83, doc-parse ≥ 90 %, schedule adherence ≥ 95 %

Solution Delivered

Event-Driven Secure Ingestion

  • S3 landing zone (SHA-256) → Step Functions OCR orchestration
  • Kafka Connect streams; KEDA auto-scales GPU pods
  • Tenant-scoped KMS keys ensure PII isolation

GenAI Document Intelligence

  • Tesseract + LayoutLMv3: parse accuracy 72 % → 93 %
  • LangChain embeddings (bge-base-en) → pgvector; retrieval < 200 ms

Feature Store & Model-Ops

  • Feast online/offline store, versioned via Git tags
  • MLflow CI/CD on EKS; SHAP explainability PDF for regulators

Risk-Scoring Micro-Services

  • Go + gRPC; AB testing via header flag
  • p95 latency 120 ms; AUC 0.84

Compliance & Observability

  • CloudTrail, GuardDuty, AWS Config → Panoptica dashboard
  • Evidence auto-harvested; audit prep effort –80 h

Cost & Scale Optimisation

  • Spot GPU Fargate for embeddings (–35 % spend)
  • Aurora auto-pause during lull; infra / app cost ratio 0.18

Execution Journey (24 Weeks, 11 Sprints)

Sprint Phase
Key Deliverables
KPI Shift
Predictability
Sprints 0
Discovery, threat model
OCR baseline 72 %
100 % gate pass
Sprints 1
S3 ingest, Kafka PoC
P95 ingest 5.4 s → 1.8 s
Risk-WIP 14 %
Sprints 2
GPU OCR workers
Doc accuracy 72 → 88 %
Buffer unused
Sprints 3
Embedding pipeline
Retrieval 400 → 190 ms
Flex GenAI 24 h
Sprints 4
Feature store + ML CI/CD
AUC 0.72 → 0.79
0 Slip
Sprints 5
gRPC risk service
AUC 0.79 → 0.83
Hot-fix 0
Sprints 6
Explainability, compliance scripts
Audit evidence 60 %
Budget +4 %
Sprints 7
5 × load, GPU optimiser
GPU cost –35 %
0 Slip
Sprints 8
SOC 2 dry-run
Doc 93 %
Flex security 16 h
Sprints 9
Partner UAT
SLA 5 d → 46 h
0 Slip
Sprints 10
Bank pilot drills
SLA 46 → 38 h
0 Slip
Sprints 11
Prod launch
0-day slip
Delivered 1 day early

Buffer engineer filled a leave gap in 4 h → velocity dip 0 SP.

Business Impact

Metric
Before
After
Delta
Underwriting Turnaround
5 days
2.5 days
–50 %
Model AUC
0.72
0.84
+12 pp
Doc-Parse Accuracy
72 %
93 %
+21 pp
SOC 2 Status
N/A
Passed 3 mo early
–80 h audit prep
Infra Cost (GPU)
Baseline
–35 %
–35 %
New ARR
0
US $ 7.8 M
+7.8 M

Predictability premium (≈ 9 % blended rate) paid back in one sprint by avoiding a four-week slip valued at ~US $ 1.1 M.

Why Steady Rabbit?

Core-Flex Micro-GCC

SMEs parachuted in 48 h; Buffer absorbed PTO risk.

SteadCAST Predictability

Live risk & capacity analytics → 97 % sprint adherence.

Shift-Left Gates

7 Plan-Left steps cut rework 42 %.

AI + Compliance Depth

LLM pipelines, SHAP explainability, SOC 2 evidence baked into DevSecOps.

Outcome-Linked Fees

KPIs (AUC, SLA, ARR) tied to squad incentives.

Transparent Partnership

Weekly C-suite demos, shared burn charts, zero surprises.

Client Testimonial

Steady Rabbit

CTO & Co-Founder

Confidential FinTech Scale-up

Steady Rabbit delivered the impossible—an AI-powered, audit-ready underwriting engine in half a year. Our lenders and investors were thrilled, and we never burned a weekend. The Micro-GCC model is now our secret weapon