The Challenge

• ~20 years of business logic grown into > 1500 stored procs
• High interdependence
• Difficult to debug and maintain
• Difficult to test
• Challenging to integrate into CI/CD pipeline with automated testing

Large Language Models

What They Are

• Transformer-based neural networks trained on massive text corpora
• Capable of understanding and generating human-like text
• Foundation for modern AI assistants and code generation

Evolution of Context Windows

Frontier Model Architecture

Not just an LLM - a complex stack of tools

Our Migration Pipeline

One-Shot Code Migration Prompt

"You need to migrate SQL to C# using Entity Framework Core..."

"Stored procs in MS SQL, output in C# in this directory <...>, 
use this namespace <...>, DB Contexts are CrmDbContext,..., 
use DI for dependencies, create Interface as well..."

"SP: fma_web_Processors_Tier_Check, Tables:..., EF names:..."

One-Shot Code Migration (cont'd)

public class Functions
{    
    public async Task<int> CustomerHasFlag(string customerID, int categoryId)
    {
        ...

public class fma_web_Limits_Check : Ifma_web_Limits_Check
{
    private readonly CrmDbContext _crmDbContext;
    ...

Highest priority is perfect match of logic flow. 
Be aware of different null handling...

Full Migration Prompt

One-Shot Test Generation

IF @CustomerID IN ('511555') --AND @SiteID = 1
BEGIN
    SET @returnProcessorCode = 'Compsta3D'
    SELECT @returnProcessorCode
    RETURN
END

// Special customer 511555 (should return Compsta3D)
testParams.Add(new TestParamsAndResult()
{
    ParameterValues = new List<object>()
    {
        "511555", // CustomerID
        "visa",   // CardType
        100.00m,  // Amount
        1,        // SiteID
        "I",      // LocationCode
        CreditCardNumberGenerator.GetNumberByType("visa", random) // CreditCardNo
    }
});

Focus: Hit all code paths without superfluous test cases

Migration UI

Browser-based dashboard

• Progress tracking of migration status
• Historical tracking of Pass/Fail
• Manual triggering of individual steps in the pipeline

Common Migration Issues

• LEFT JOIN vs .Join(): Different null handling semantics
• Casing: IsIBO vs. IsIbo inconsistencies
• Ambiguous References: TransactionStatusCode without alias
• DateTime.Now: Time zone and testability issues

✨ Core logic is understood and translated well!

One-Shot vs. Agentic Approaches

Automating with Claude Code

The Feedback Pipeline

• Each pipeline step callable from outside
• Investigative tools for debugging
• Hands-off execution
• Human-in-the-loop only at the end

Claude Code Functions

Via dotnet run command line

Learnings, Tips & Tricks

• Things are moving fast!
• Massive Context: Unlike 2024, use it all!
• Be Concise: Understand deeply, constrain tightly
• Fall Forward: Design prompts for success paths
• Classic Techniques Still Work:
  • Roleplaying: "You are a world-class engineer...", encouragement, bribery
• Read the Guides: OpenAI, Anthropic, Google prompt guides remain valuable

Resources & Further Reading

Prompting Guides

• OpenAI Prompt Engineering Guide
• Anthropic Prompt Engineering
• Claude Code Getting Started

AI & Future

• Bond Capital - The AI Report

This Presentation

• GitHub Repository

GitHub QR Code

Scan for GitHub repo

Questions?

Thank you!