A three-build progression on the UiPath platform that traces an enterprise automation strategy from foundations to synthesis: from deterministic loan intake, through agentic claims adjudication with human-in-the-loop review, to a unified regulator-grade decisioning platform with full audit and disclosure.
Meridian Financial Group is a fictional $40B-AUM regional bank that has just acquired Cascade Health Partners, a specialty insurance subsidiary. The merger forces three immediate questions onto the automation team:
- Operational — how do you consolidate 2,000+ back-office processes across loan origination, claims adjudication, and compliance reporting without doubling headcount?
- Regulatory — how do you make every customer-facing decision defensible to an auditor inside a 90-day post-merger window?
- Strategic — how do you become "agentic-first" inside twelve months and still hold the line on governance?
The three builds in this repository are a study in answering those questions progressively — not with hype, and not by replacing existing work with new work, but by layering capability so each build keeps the discipline of the one before it.
Deterministic RPA. No agents. A production-grade dispatcher/performer pair on classic UiPath Studio + Orchestrator, built on REFramework, that ingests loan applications from a shared inbox, validates them, scores them against business rules, and routes to one of five outcomes. The foundational discipline: object-oriented design, named exception classes, full audit on Orchestrator queue items.
Before anyone trusts you with agents, you have to prove you can still do the boring stuff brilliantly.
Agentic capability under enterprise constraints. Two LLM agents orchestrated by a Maestro BPMN flow. The first uses Document Understanding as a Tool to extract structured data from CMS-1500 PDFs. The second adjudicates under five hard rules, with Anthropic's Claude available as a second-opinion Tool for ambiguous medical-necessity calls. High-dollar reviews suspend the Maestro instance in Action Center for human review — a runtime model meaningfully different from run-to-completion RPA.
Agents handle the cases their inputs are trustworthy for; deterministic fallbacks handle the cases that don't meet the agent's preconditions.
The synthesis. A single agentic pipeline that serves both lines of
business — loan applications and insurance claims — through a unified
classify → intake → risk → compliance → disclosure flow. Risk
decisioning is a hybrid: deterministic for loans (a real-bank credit
ladder), agentic for claims (Week 2's pattern, reused). Every stage
appends to an immutable decisionEvent record that, by end of flow,
contains the complete ordered trace of how the disposition was reached
— so an examiner can reconstruct any decision in a single record
without log archeology.
Governance is the product of doing automation well, not the constraint on it.
| Property | Week 1 | Week 2 | Week 3 |
|---|---|---|---|
| Decision-making layer | Rule-based scripts | LLM agents | LLM agents + deterministic |
| Orchestration | Orchestrator queues | Maestro BPMN | Maestro BPMN |
| Human-in-the-loop | Action Center form (>$500K) | Action Center, suspend-and-wait (>$10K REVIEW) | Action Center, suspend-and-wait (escalation_required) |
| External integrations | Gmail, core-banking VBO (stub) | Gmail, Sheets, Claude, DU | Gmail, Claude, DU, PDFMonkey, Storage Bucket |
| Output | Routing decision on queue | Audit log row in Google Sheets | PDF disclosure letter + email + bucket archive |
| Governance posture | Exception classes + Orchestrator audit | Per-agent contract discipline | Formal Model Card + Decision Logging Spec + Audit Retention Matrix + Override Audit Trail Standard |
| Lines of business | Loans only | Claims only | Loans + claims, unified |
Each build adds capability without abandoning the prior build's
discipline. Week 1's exception classes survive into Week 3's
business-exception path. Week 2's confidence-driven branching pattern
survives into Week 3's classifier-confidence gate. Week 3's
decisionEvent record is the production-grade evolution of Week 2's
Google Sheet audit log.
MeridianBank_Project/
├── README.md ← you are here
├── .gitignore macOS / Windows / UiPath build cache
│
├── Week_1_LoanIntake/ Build 1 — deterministic RPA
│ ├── README.md
│ ├── docs/ SDD + architecture diagram
│ ├── source/ LoanIntake.Loader / .Worker / .Library
│ └── test-data/ sample applications + 500-email load harness
│
├── Week_2_CascadeHealth/ Build 2 — agentic with HITL
│ ├── README.md
│ ├── docs/ SDD, governance packet, data scheme, phase findings
│ ├── source/ 6-project solution (intake + adjudication agents,
│ │ DU project, Maestro BPMN, bucket RPA, Action Center app)
│ └── test-data/ 40 CMS-1500 PDFs + deterministic generator
│
└── Week_3_RDD/ Build 3 — synthesis with full governance
├── README.md
├── docs/ board memo, governance packet, architecture diagram
├── source/ 8-project solution (4 LLM agents + Maestro BPMN +
│ Action Center app + 2 RPA bucket bridges)
└── test-data/ 50 self-labeling JSON payloads + generator
Built entirely on the UiPath platform across:
- UiPath Studio (classic, Windows) — Week 1 dispatcher/performer/library
- UiPath Studio Web — Weeks 2 and 3 solution composition
- UiPath Maestro — BPMN orchestration in Weeks 2 and 3
- UiPath Agent Builder — LLM agent definition for Weeks 2 and 3
- UiPath Document Understanding (Modern) — Week 2 CMS-1500 extraction
- UiPath Action Center — human-in-the-loop in all three builds
- UiPath Orchestrator — queues, assets, storage buckets, schedules across all three
- UiPath Integration Service — connectors for Gmail, Sheets, Claude, PDFMonkey
If you have ten minutes and want the shortest path to evaluating this work, read in this order:
- Week 3 README — the synthesis and the most fully-realized build
- Week 3 board memo — the strategic case the build defends
- Week 2 governance packet — the honest version of what's hard about agentic systems
- Week 1 README — for the foundational discipline that underpins the rest
If you have an hour and want regulator-grade detail, add the Week 3 Governance Packet — Model Card + Decision Logging Specification + Audit Retention Matrix + Human Override Audit Trail Standard.
Christopher Williams · c.williams101198@gmail.com
This is a portfolio build. All companies, customers, claims, and
applicants are fictional. Connection definitions in the repo use
AuthenticateAfterDeployment — no credentials are committed.