AI-Led Modernization,
Proven at Every Engineering Gate
Challenges
An insurance provider's core platform had run on HP NonStop for nearly three decades: 1.45 million lines of code across 617 COBOL programs, a 182-screen C++ GUI, and .NET components, orchestrated by 546 batch schedules and 293 database tables.
Documentation had drifted far behind the live system, while deep platform knowledge sat with a handful of long-tenured experts. The challenge was to reverse-engineer this tightly coupled estate, preserve every business rule, and prove a modern .NET and Azure path without disrupting live insurance clients.
Solutions
AI-Led Modernization, Proven at Every Engineering Gate
Zensar delivered a 16-working-day proof of concept (PoC) built on a Claude-powered, human-validated modernization model. Claude served as the core generative AI engine - interpreting legacy code, reconstructing business logic, generating modernization artifacts, and supporting validation.
A semantic model pipeline converted 24 COBOL and 2 C compiled listings into verified knowledge models, call graphs, dependency views, business rules, and data mappings, while nine catalogue agents and three purpose-built agents accelerated discovery, code generation, gap analysis, and validation.
The team forward-engineered the workload into four .NET 8 projects on Azure and implemented 171 Microsoft RulesEngine rules. A two-stream validation approach and five-level equivalence framework reconciled six output tables against an independent golden reference - turning every success criterion into measurable evidence for sign-off.
Feature 1:
A semantic model pipeline distilled ~84K lines across 26 compiled listings into schema-validated knowledge models, call graphs, dependency views, and a domain map - treating compiled artifacts, not program names, as the source of truth.
Feature 2:
Nine catalogue agents drove discovery, architecture, build, and validation, with Claude powering the reasoning behind each. Three purpose-built agents - Code Gap Analyzer, Code Agent, and Validation Agent - closed generation gaps and checked business-rule projections, routing low-confidence outputs to specialist review. Claude's speed met human control at every critical gate.
Feature 3:
The legacy process was rebuilt as four .NET 8 projects on Azure SQL, Storage, App Insights, and Microsoft RulesEngine - generating 18,289 lines of production C# and 17,692 lines of test code, backed by 171 property-tested rules.
Feature 4:
A five-level framework verified row presence, key integrity, field values, derived logic, and cross-table relationships. The 2,280-record input populated six output tables with 100% field match across non-excluded columns and zero open defects.
Solution enablers
Tech Stack
- HP NonStop, COBOL, C, TAL, SQL/MP, and Enscribe
- .NET 8, C#, Azure SQL, Azure Storage, and App Insights
- Microsoft RulesEngine and xUnit
- Semantic Model Pipeline and Connected Intelligence corpus
- Claude as the core generative AI and reasoning engine for code understanding, artifact generation, and validation support
- Claude-powered AI agents for legacy parsing, rule extraction, code generation, gap analysis and validation
Zensar Services
- AI-led legacy discovery and reverse-engineering
- Business-rule extraction and requirements reconstruction
- Cloud-native application modernization
- Equivalence testing, reconciliation, and evidence-based governance
Business outcomes
Proving complex legacy modernization with measurable equivalence
The PoC proved that a complex HP NonStop workload could be successfully modernized to .NET 8 and Azure while maintaining 100% functional and business-field equivalence. By leveraging Claude for legacy code analysis, business-rule reconstruction, code generation, gap analysis, and validation, alongside specialist oversight and a five-level reconciliation framework, Zensar transformed approximately 84,000 lines of multi-language legacy code into a cloud-ready .NET 8 solution in just 16 working days. The engagement delivered zero open defects, accelerated modernization efforts, and established a reusable foundation for future large-scale legacy transformation initiatives.
Feature 1:
6 of 6 success criteria and all four named gates achieved
Feature 2:
100% field match across non-excluded columns; zero open defects
Feature 3:
26 compiled listings decoded and 2,280 records processed
Feature 4:
12 custom agents