Loan Origination Pipeline for a Bank

🏦 Project: Loan Origination Pipeline for a Bank

📌 Context (as-is state)

We started by working on a legacy loan process system in a bank.

The system was a classic monolith that combined:

• loan application intake
• client verification
• collateral processing
• credit dossier assembly
• deal origination
• integrations with external services (Credit Bureau, collateral registries, CRM, agent's workstation)

System problems:

• tight coupling of modules within the monolith
• long application processing time
• lack of transparent stage control
• frequent failures during integrations
• inability to scale individual stages (e.g., collateral verification)
• difficulty implementing new rules and products

In essence, it was a "large data pipeline" that failed with any instability in one of its components.

---

🎯 Transformation Goal

We set the following goals:

• decompose the monolith into independent domain modules
• provide a managed business process orchestrator
• enhance the stability of the loan origination pipeline
• implement AI-powered document processing automation
• make the system scalable at each stage

---

🧩 Domain Decomposition

We divided the system into independent domain boundaries:

1. 📥 Application Intake

• a single entry point for applications (CRM / partners / agent's workstation)
• data normalization
• primary validation
• application deduplication

---

2. 🧾 Client and Collateral Check (Risk & Collateral Check)

• integrations with external sources (Credit Bureau, registries, anti-fraud)
• client scoring
• collateral property valuation
• calculation of limits and conditions

---

3. 🤖 Dossier Assembly using AI (Document Intelligence)

One of the key transformation modules.

We implemented an AI pipeline:

• document upload (PDFs, images, scans)

• OCR and entity extraction

• LLM processing for:

• document classification

• extraction of key fields (Full Name, amounts, dates, collateral objects)

• completeness check of the dossier

• automated generation of a structured credit dossier

Result:

• reduction in manual document review
• decrease in errors during dossier assembly
• acceleration of the deal preparation stage

---

4. 📦 Case Assembly Module

• collection of all verification results
• aggregation of data from different domains
• construction of a unified credit case
• data completeness control before the final decision

---

5. ✍️ Deal Origination

• generation of contracts and loan documents
• integration with EDMS (Electronic Document Management System)
• final agreement on terms
• recording the transaction in the bank's accounting systems

---

⚙️ New Solution Architecture

🧠 Orchestration: Temporal

We used Temporal as the core process management engine:

• each loan = a separate workflow

• each stage = an activity

• guarantees:

• retry on errors

• state preservation

• recovery after failures

• idempotency of operations

Temporal allowed us to transform the loan process into an observable state machine, rather than "a script that sometimes crashes".

---

☸️ Infrastructure: Kubernetes

The entire system was deployed on Kubernetes:

• each domain module - a separate deployment

• horizontal scaling:

• OCR/AI separately

• risk scoring separately

• integrations separately

• fault isolation between services

---

🧱 Microservices Architecture

Each domain became a separate service:

• Application Service
• Risk Service
• Collateral Service
• Document AI Service
• Case Assembly Service
• Deal Service

Communications:

• synchronous calls (gRPC/HTTP) for quick verifications
• asynchronous events via queues for heavy operations

---

🔄 How the Process Looks (end-to-end)

1. A loan application is submitted
2. Temporal initiates a workflow
3. In parallel:

• the client is verified
• collateral is evaluated
• AI document processing is initiated

4. Data flows into Case Assembly
5. Completeness and consistency are checked
6. If everything is OK - deal origination is initiated
7. The transaction is recorded in the bank's systems

---

🚀 Key Results

After transitioning to the new architecture:

• application processing time significantly decreased
• process stability increased (no "pipeline crashes")
• individual stages can be scaled independently
• implementation of new loan products accelerated
• full traceability for each loan was achieved
• manual document processing workload significantly reduced

---

💡 Main Architectural Shift

The most important change was not technological, but conceptual:

we stopped thinking about a loan as a monolith and started thinking about it as a managed distributed process (workflow), where each step is independent, observable, and recoverable.