Verified Diploma
A large federal financial agency needed to modernize paper-heavy operations without disrupting a 60-year-old master database. I built an AI-assisted translation layer over PEGA's agile delivery methodology, converting raw stakeholder transcripts and legacy processing rules directly into structured user stories, UAT scripts, and PEGA Blueprint architectures. 150+ stories shipped. 40% documentation acceleration. Zero legacy database disruptions.
The agency processes trillions of dollars annually but was running on fragmented, paper-heavy workflows tied to a 60-year-old IT architecture. Leadership had a mandate to modernize, but a traditional system replacement carried catastrophic risk. Disrupting the legacy database was not an option.
The solution was PEGA's "Digital Agility Layer" · a low-code approach that wraps modern digital workflows around the existing architecture without touching the core database. But modernization projects fail at the requirements phase, not the build phase. Converting complex federal processing rules from Direct Capture of Objectives (DCO) sessions into actionable PEGA architectures manually was a bottleneck that was grinding development sprints to a halt.
The requirements translation phase was fully manual. A human analyst had to read raw DCO session transcripts, extract processing logic, write SAFe-compliant user stories, and map them to PEGA architectures · one by one. Errors at this stage contaminated downstream development. SLA rules that were wrong meant non-compliant case routing. That could not happen in a federal financial context.
I was brought in as the bridge between federal stakeholders and the PEGA delivery team. My job was to fix the translation layer.
The first prompt versions produced generic agile output. They missed PEGA-specific persona mappings entirely and failed to capture the SLA escalation rules embedded in federal processing language. The AI was treating it like a standard product backlog exercise. It wasn't.
The jump from V1 to V4 required three key additions: a federal architect role context, strict negative constraints on SLA generation (extract only, never invent), and a forced UAT traceability requirement linking every test script back to its parent user story ID. Without those three constraints, the output was unusable in a compliance environment.
Operating in a FedRAMP-compliant environment meant one thing: invented outputs have legal consequences. I built three hard constraints into every production prompt.
1. SLA Extraction Lock: AI is strictly prohibited from generating or estimating Service Level Agreement timelines. All routing logic must be deterministically extracted from the federal source text · no inference, no estimation.
2. UAT Traceability: Every generated UAT script must cross-reference its parent user story ID. No orphaned test scripts. Audit-ready output by default.
3. SME Judgment Gate: All case lifecycles, personas, and data interfaces go through human review before entering PEGA Agile Studio. I was the human-in-the-loop on every output.
The SLA lock was the most important rule. In federal case management, a fabricated SLA timeline does not just create a bad user story · it creates a non-compliant workflow that reaches production. That is a different category of problem.
| Evaluation Criterion | V1 Output | V4 Production Output | Status |
|---|---|---|---|
| PEGA Microjourney Alignment (SME Review) | 20% | 100% | PASS |
| SLA Rule Fidelity (Source-to-Output Match) | 45% · active hallucinations | 100% strict extraction | PASS |
| UAT Script Traceability (ID Cross-Reference) | 0% | 100% | PASS |
| Documentation & Backlog Acceleration | Baseline | 40% faster time-to-groom | PASS |
V1's SLA hallucination failure was the clearest example of why the guardrails existed. The AI generated SLA timelines that were not present in the source text · plausible-sounding but non-compliant. Adding the strict extraction constraint and running a source-to-output match against the federal processing rules eliminated the problem entirely. The AI handles translation volume. I handle accuracy against compliance standards.
The translation layer I built was scoped to one modernization workstream. Applied across all Microjourney tracks in a full PEGA enterprise deployment, the same AI-assisted methodology would compress the entire Discover and Prepare phase · the most time-intensive part of any low-code modernization · by a comparable margin without additional headcount.