A global fintech leader ran its [i]payments reconciliation[/i] on a legacy platform their team could no longer build against.

A global fintech leader ran its [i]payments reconciliation[/i] on a legacy platform their team could no longer build against.

A global fintech leader ran its [i]payments reconciliation[/i] on a legacy platform their team could no longer build against.

A global fintech leader ran its [i]payments reconciliation[/i] on a legacy platform their team could no longer build against.
A global fintech leader ran its [i]payments reconciliation[/i] on a legacy platform their team could no longer build against.

The platform, written in Delphi, managed exception ticketing, executed funds movement, and carried seven-year data retention obligations. A single drive failure or bad reboot created a real probability the system would never come back up, and manual reconciliation at that volume was not an option.

The Delphi compiler license had lapsed

The Delphi compiler license had lapsed

The Delphi compiler license had lapsed

The Delphi compiler license had lapsed
The Delphi compiler license had lapsed

The compiled language can no longer be licensed. Rebuilding the system was not an option without the tools to compile it.

End of life operating system

End of life operating system

End of life operating system

End of life operating system
End of life operating system

The OS running the platform was no longer receiving security patches or vendor support.

Discontinued hardware

Discontinued hardware

Discontinued hardware

Discontinued hardware
Discontinued hardware

The physical machine in the data center could not be replaced. A drive failure with no path to
recovery.

Institutional knowledge gone

Institutional knowledge gone

Institutional knowledge gone

Institutional knowledge gone
Institutional knowledge gone

Every engineer who built the system
had left. No documentation existed outside the source code itself.

What Praxent Did

PraxentDecode[sup]SM[/sup] ran an agentic reverse-engineering process against the full codebase.

PraxentDecode[sup]SM[/sup] ran an agentic reverse-engineering process against the full codebase.

PraxentDecode[sup]SM[/sup] ran an agentic reverse-engineering process against the full codebase.

PraxentDecode[sup]SM[/sup] ran an agentic reverse-engineering process against the full codebase.
PraxentDecode[sup]SM[/sup] ran an agentic reverse-engineering process against the full codebase.

Inventory and
classification

Inventory and
classification

Inventory and
classification

Inventory and
classification
Inventory and
classification

Ever y file fingerprinted and scored on a six-factor complexity model.

Structure analysis

Structure analysis

Structure analysis

Structure analysis
Structure analysis

Blueprint of every class, form, API surface, and module.

Dependency analysis

Dependency analysis

Dependency analysis

Dependency analysis
Dependency analysis

Full dependency graph with circular-dependency and coupling-risk reports.

Data model documentation

Data model documentation

Data model documentation

Data model documentation
Data model documentation

Every record, class, and database schema mapped with code-to-SQL traceability.

Business logic extraction

Business logic extraction

Business logic extraction

Business logic extraction
Business logic extraction

Calculations, validations, and
SQL intent recovered line by line. 143 SME questions raised and prioritized.

Draft specification synthesis

Draft specification synthesis

Draft specification synthesis

Draft specification synthesis
Draft specification synthesis

Findings organized across seven business domains with traced requirements and user stories.

Final specification assembly

Final specification assembly

Final specification assembly

Final specification assembly
Final specification assembly

SME answers folded back in, producing implementation-ready requirements with full traceability.

Requirements
validation

Requirements
validation

Requirements
validation

Requirements
validation
Requirements
validation

Cross-checked against source code, database structures, and documented system behavior.

PraxentDecode℠ often surfaces architectural boundaries the original team did not know existed. In this engagement what the organization treated as one system was actually three applications running as one: downloads, daily reconciliation, and bank reconciliation.
Winding mountain road illuminated by car headlights at night, surrounded by rocky cliffs and dense forest with fog filling the distant valley.

Every phase of Decode has a Praxent forward-deployed fintech engineer in the loop. They calibrate scope, interpret findings against payments patterns, decide which of the SME questions need human input, and translate the recovered logic into a build plan compliance can sign off on. That loop produces specs you can act on.

Paul Tidwell
Head of AI
,
Praxent
Portrait of a middle-aged man with short brown hair and a trimmed beard, wearing a light blue dress shirt and a dark suit jacket.
What Decode Produces
A reverse-engineering agent reads the source code, builds an internal model of how the system behaves, raises the questions only humans can answer, and folds those answers back into traceable specs.

EXAMPLE BUSINESS RULE EXTRACTED

The bank statement balance is calculated as the sum of all line item amounts for the statement period, but only for top-level items where the parent line item is null or 0. Detail and child line items are excluded to prevent double-counting.

Rules like this live inside stored procedures and calculation methods, often
hard-coded hundreds of lines deep. They are the kind of detail a rebuild misses unless the original logic is recovered first.

Legacy modernizations break in the third month of production when a business rule no one remembered surfaces incorrectly. Decode pulls those rules out of the source first. SMEs validate them before the build starts. Parity comes faster, with less risk.

Paul Tidwell
Head of AI
,
Praxent
Portrait of a middle-aged man with short brown hair and a trimmed beard, wearing a light blue dress shirt and a dark suit jacket.
The Output
14K+
Lines of legacy code mapped
714
Business requirements recovered
143
SME questions generated
7
Business domains mapped
5
Security vulnerabilities surfaced
Security Findings
White lightning bolt icon with sharp angles.

Hardcoded credentials

Hardcoded credentials

Hardcoded credentials

Hardcoded credentials
Hardcoded credentials

stored in source, visible to any developer with repository access.

White lightning bolt icon with sharp angles.

SQL injection attack vectors

SQL injection attack vectors

SQL injection attack vectors

SQL injection attack vectors
SQL injection attack vectors

stored in source, visible to any developer with repository access.

White lightning bolt icon with sharp angles.

Buffer overrun vulnerabilities

Buffer overrun vulnerabilities

Buffer overrun vulnerabilities

Buffer overrun vulnerabilities
Buffer overrun vulnerabilities

stored in source, visible to any developer with repository access.

White lightning bolt icon with sharp angles.

Two embedded Y2K pivot dates

Two embedded Y2K pivot dates

Two embedded Y2K pivot dates

Two embedded Y2K pivot dates
Two embedded Y2K pivot dates

stored in source, visible to any developer with repository access.

White lightning bolt icon with sharp angles.

Undocumented exception paths

Undocumented exception paths

Undocumented exception paths

Undocumented exception paths
Undocumented exception paths

stored in source, visible to any developer with repository access.

Modern security vulnerabilities, exploited by AI-powered attack tools, are making 20-year-old codebases even more exposed. The risk is getting higher. And the barriers are getting lower.

Paul Tidwell
,
Head of AI, Praxent
Portrait of a middle-aged man with short brown hair and a trimmed beard, wearing a light blue dress shirt and a dark suit jacket.
The Results

PraxentDecode[sup]SM[/sup] ran an agentic reverse-engineering process against the full codebase.

PraxentDecode[sup]SM[/sup] ran an agentic reverse-engineering process against the full codebase.

PraxentDecode[sup]SM[/sup] ran an agentic reverse-engineering process against the full codebase.

PraxentDecode[sup]SM[/sup] ran an agentic reverse-engineering process against the full codebase.
PraxentDecode[sup]SM[/sup] ran an agentic reverse-engineering process against the full codebase.
White lightning bolt icon with sharp angles.

Hardcoded credentials

Hardcoded credentials

Hardcoded credentials

Hardcoded credentials
Hardcoded credentials

stored in source, visible to any developer with repository access.

White lightning bolt icon with sharp angles.

SQL injection attack vectors

SQL injection attack vectors

SQL injection attack vectors

SQL injection attack vectors
SQL injection attack vectors

stored in source, visible to any developer with repository access.

White lightning bolt icon with sharp angles.

Buffer overrun vulnerabilities

Buffer overrun vulnerabilities

Buffer overrun vulnerabilities

Buffer overrun vulnerabilities
Buffer overrun vulnerabilities

stored in source, visible to any developer with repository access.

White lightning bolt icon with sharp angles.

Two embedded Y2K pivot dates

Two embedded Y2K pivot dates

Two embedded Y2K pivot dates

Two embedded Y2K pivot dates
Two embedded Y2K pivot dates

stored in source, visible to any developer with repository access.

White lightning bolt icon with sharp angles.

Undocumented exception paths

Undocumented exception paths

Undocumented exception paths

Undocumented exception paths
Undocumented exception paths

stored in source, visible to any developer with repository access.

Praxent's financial-services engineers run a parallel test harness against recovered specs, validating every business rule before it reaches production.
SYSTEM PARITY GUARANTEE

Two validation passes get Decode from 95% to 100% parity.

Two validation passes get Decode from 95% to 100% parity.

Two validation passes get Decode from 95% to 100% parity.

Two validation passes get Decode from 95% to 100% parity.
Two validation passes get Decode from 95% to 100% parity.
VALIDATION PASS 1

Independent scoring

Pre-implementation

  1. 1
    Praxent generates the analysis
  2. 2
    Independent model reads recovered requirements
  3. 3
    Locates each requirement in source, scores coverage
  4. 4
    Validation pass runs 2–3 different ways for the same engagement
VALIDATION PASS 2

Parallel test harness

Implementation phase

SAME INPUT
OLD SYSTEM Legacy platform
NEW SYSTEM Rebuilt platform
SAME OUTPUT EXPECTED

Tested against synthetic, real, and historical data. Differences become test failures.

Before LLMs, this kind of recovery relied on people, and the error rate was significantly higher than 5%. The structured, validated, agentic approach is faster and more accurate than what a human team typically produced.

We're in an era where you can fully understand and rebuild parts of your system that are slowing you down. When you can do that for a fraction of the cost, you fundamentally reevaluate what legacy means and how it is affecting your velocity as a business.

Paul Tidwell
,
Head of AI, Praxent
Portrait of a middle-aged man with short brown hair and a trimmed beard, wearing a light blue dress shirt and a dark suit jacket.

Traditional Manual Discovery Approach

Traditional Manual Discovery Approach

Traditional Manual Discovery Approach

Traditional Manual Discovery Approach
Traditional Manual Discovery Approach
~6
months
~3,840
person-hours
~$250K
in engineering time
40-60
SME requirements workshops

PraxentDecodeSM

Text saying 'Powered by Claude' with an orange asterisk-like symbol between the words.
~2
weeks
100
hours of runtime
$2.5K
in tokens
-3
SME requirements workshops