GitHub - scalefirstai/EvolutionAI: Reference Implementation for Large Scale Modernization Effort in the Banking Domain using AI

5 min read Original article ↗

AI Agentic Architecture for Legacy Banking Modernization

License CI Java Quarkus

What This Project Demonstrates

EvolutionAI is a production-grade reference implementation showing how multi-agent AI orchestration can modernize legacy COBOL banking systems into cloud-native Java/Quarkus services — with full observability, compliance verification, and executive visibility.

  • Source: AWS CardDemo — a COBOL/CICS credit card management application
  • Target: Apache Fineract-compatible REST APIs built on Java 17 / Quarkus
  • Method: Six specialized AI agents working in a coordinated pipeline with human-in-the-loop design decisions
  • Result: Browsable conversion artifacts, a living metrics dashboard, and a repeatable methodology

Key Results

Metric Traditional AI-Driven (This Project)
Cost per COBOL line $15 — $25 < $2.50
Human hours per module 200+ < 40
Timeline (full migration) 3 — 5 years 6 — 12 months
Behavioral equivalence Manual testing 99.7% automated pass rate
Compliance verification Post-hoc audit Continuous, per-module

Quick Start

# Clone the repository
git clone https://github.com/scalefirstai/EvolutionAI.git
cd EvolutionAI

# Set up COBOL source (clones AWS CardDemo)
./scripts/setup.sh

# Launch all services (dashboard, observability, metrics)
docker-compose up

# Open the dashboard
open http://localhost:3000

Service URLs

Service URL Purpose
Dashboard http://localhost:3000 Executive visibility and metrics
Langfuse http://localhost:3100 LLM trace inspection
Prometheus http://localhost:9090 Time-series metrics
Grafana http://localhost:3001 Operational dashboards

Architecture

                    ┌─────────────────────────────────────────────┐
                    │              Mayor (Orchestrator)            │
                    │     Decomposes work, manages dependencies    │
                    └────────┬──────────┬──────────┬──────────────┘
                             │          │          │
                    ┌────────▼───┐ ┌────▼────┐ ┌──▼──────────┐
                    │  Analyst   │ │ Polecat │ │  Polecat    │
                    │  (Reverse  │ │ (Trans- │ │  (Trans-    │
                    │   Engineer)│ │  late)  │ │   late)     │
                    └────────┬───┘ └────┬────┘ └──┬──────────┘
                             │          │          │
                    ┌────────▼──────────▼──────────▼──────────────┐
                    │              Witness (Quality Gate)          │
                    │   Behavioral equivalence, security, FINOS   │
                    └────────────────────┬────────────────────────┘
                                         │
                    ┌────────────────────▼────────────────────────┐
                    │            Refinery (Merge Manager)          │
                    │    Branch hygiene, conflict resolution, PRs  │
                    └─────────────────────────────────────────────┘
                                         │
                    ┌────────────────────▼────────────────────────┐
                    │             Deacon (Watchdog)                │
                    │   Health monitoring, cost tracking, alerts   │
                    └─────────────────────────────────────────────┘

Agent Roles

Agent Model Responsibility
Mayor Opus Orchestrates work, decomposes tasks, manages dependencies
Analyst Sonnet Reverse-engineers COBOL modules, extracts business rules
Polecat Sonnet Translates COBOL to Java/Quarkus (runs in parallel)
Witness Opus Reviews for quality, security, compliance — has veto power
Refinery Sonnet Manages merge queue, resolves conflicts
Deacon Haiku Monitors fleet health, tracks costs, fires alerts

The Five-Stage Loop

Every COBOL module flows through a repeatable five-stage pipeline:

  ┌──────────┐    ┌──────────┐    ┌───────────┐    ┌──────────┐    ┌──────────┐
  │ DISCOVER │───▶│  DESIGN  │───▶│ TRANSLATE │───▶│ VALIDATE │───▶│  DEPLOY  │
  │ (AI)     │    │ (Human)  │    │ (AI)      │    │ (AI)     │    │ (AI)     │
  └──────────┘    └──────────┘    └───────────┘    └──────────┘    └──────────┘
   Analyst         Architect       Polecat          Witness         Refinery
   agent           review         agent(s)          agent           agent
  1. Discover — Analyst agent maps COBOL structure, extracts business rules, documents dependencies
  2. Design — Human architect reviews analysis, makes target architecture decisions
  3. Translate — Polecat agents convert COBOL to Java/Quarkus with tests and OpenAPI specs
  4. Validate — Witness agent runs equivalence tests, checks security, verifies FINOS CDM compliance
  5. Deploy — Refinery merges approved code; Deacon monitors post-deploy metrics

Demo Conversion: CardDemo Modules

Module Business Function Target API Status
COSGN00C User authentication / login /api/v1/auth/login Converted
COCRDLIC Credit card listing /api/v1/cards Converted
COACTUPC Account update (CRUD) /api/v1/accounts/{id} Converted
COCRDUPC Credit card update /api/v1/cards/{id} Converted
COTRN00C Transaction processing /api/v1/transactions Converted
CBTRN01C Batch transaction processing Scheduled job Converted
CBACT01C Batch account processing Scheduled job Converted

Browse the Output

Each converted module includes:

Artifact Location Description
COBOL Source source/cobol/ Original COBOL program
RE Report output/docs/RE-*.md Reverse engineering analysis
Java Code output/java/ Generated Java/Quarkus implementation
Tests output/tests/ Unit and equivalence test suites
OpenAPI Spec specs/ REST API specification
Review Verdict output/docs/REVIEW-*.md Witness quality gate decision
Compliance compliance/ FINOS CDM validation report

Living Dashboard

The metrics dashboard provides four stakeholder views with real-time project visibility:

View Audience Update Frequency Key Metrics
Executive CTO, VP Eng Daily Modules completed, cost burn, ROI, risk score
Architecture Tech Leads Hourly Dependency graph, agent assignments, test coverage
Operations DevOps, SRE Real-time Agent health, token burn rate, error logs
Compliance Risk, Audit Per-event Audit trail, regulatory checks, approval chain

KPI Targets

Category Metric Target
Velocity Lines analyzed per day 50,000 — 100,000 LOC
Velocity Modules per sprint 8 — 15 modules / 2 weeks
Quality Equivalence pass rate > 99.5%
Quality First-pass code review acceptance > 85%
Cost Cost per COBOL line < $2.50
Cost Human hours per module < 40 hours
Safety Compliance check pass rate 100%
Safety Security vulnerability detection > 95%

Documentation

Document Audience Description
Executive Summary C-Suite ROI analysis, business case, risk assessment
Methodology Guide All stakeholders Five-Stage Loop explained with examples
Best Practices Engineering teams AI modernization patterns and anti-patterns
Quick Start Adopting teams Step-by-step setup and customization guide
COBOL-to-Java Mapping Developers Complete type and pattern reference
Architecture Decisions Architects ADRs for key technology choices

Open-Source Stack

Component Purpose License
AWS CardDemo Source COBOL application MIT-0
Apache Fineract Target banking API platform Apache 2.0
Quarkus Java application framework Apache 2.0
Langfuse LLM observability and tracing MIT
Prometheus Metrics collection Apache 2.0
Grafana Metrics visualization AGPL 3.0
FINOS CDM Financial compliance standard Apache 2.0

Project Structure

EvolutionAI/
├── .claude/              # AI agent definitions and skills
│   ├── agents/           # 6 specialized agent prompts
│   └── skills/           # Domain knowledge (COBOL, Fineract, observability)
├── source/cobol/         # Original COBOL source from CardDemo
├── output/
│   ├── docs/             # RE reports and witness reviews
│   ├── java/             # Generated Java/Quarkus code (buildable)
│   └── tests/            # Generated test suites
├── specs/                # OpenAPI 3.1 specifications
├── compliance/           # FINOS CDM validation reports
├── dashboard/            # Next.js metrics dashboard
├── metrics/              # Prometheus, Grafana, metrics exporter
├── docs/                 # Methodology, best practices, ADRs
├── scripts/              # Setup and demo scripts
└── docker-compose.yml    # One-command full stack startup

Contributing

See CONTRIBUTING.md for guidelines. All contributions are welcome — from documentation improvements to new module conversions.

License

This project is licensed under the Apache License 2.0 — see LICENSE for details.

Acknowledgments