AI-Assisted Billing Leakage & Contract Compliance Review
Confidential finance automation project combining a React/Supabase prototype, Python analysis pipelines, Codex subagents, and executive Excel/PDF deliverables to compare contracts, bills, payroll, support documents, and operational exports.
Confidential Context
This case study is sanitized. Client data and proprietary integrations are omitted. Work performed at Lean Tech for a confidential client. Public details are sanitized; client names, account names, internal files, and proprietary calculations are omitted.
Outcomes
- Analyzed hundreds of client/account records across contract, billing, payroll, support-document, and operational sources
- Produced executive-ready evidence workbooks and reporting packages across multiple billing opportunity categories
- Separated confirmed findings from opportunity/review pools, missing-source gaps, and human-approval lanes
- Helped shift the project from a prototype app into trusted finance and operations deliverables
Problem
The project started as a billing-leakage and contract-compliance tool. The original direction was a Lovable-built web app that compared contracts against bills, but stakeholder feedback made one thing clear: the business needed trusted analysis and evidence packs more than another polished UI.
The data was spread across contracts, bill exports, payroll/support workbooks, operational reports, PDF evidence, and manual review notes. The hard part was not only finding possible billing gaps. It was preserving the difference between confirmed findings, opportunity pools, missing-source cases, and items that still required finance or legal approval.
Approach
- Reviewed and extended the Vite/React/Supabase prototype for contract-vs-bill comparison.
- Assessed the live ingestion architecture: bill upload, contract upload, deterministic extraction, AI fallback, contract matching, discrepancy classification, versioning, and reviewer feedback loops.
- Used Python and Pandas workflows to normalize large Excel/PDF inputs and generate analysis outputs that finance stakeholders could inspect.
- Used Codex subagents for parallel review and QA over large document and workbook sets, including contract triage, payment-term matching, late-fee setup checks, and category-specific evidence review.
- Helped shape a finance-agent style backend foundation for job runs, generated artifacts, review packets, candidate workbook outputs, and future human approval decisions.
- Supported delivery operations with ClickUp-ready status updates, QA lanes, and wording that separated confirmed leakage from review inventory.
What I Delivered
- Category-level evidence workbooks for finance and operations review.
- Executive reporting packages that summarized opportunity areas and recommended action lanes.
- Contract review outputs with confidence labels, unresolved-action queues, and page/evidence references.
- Payment-terms and late-fee crosschecks that preserved weak matches instead of forcing unsafe fuzzy matches.
- Analysis guidance that helped the team move from a prototype product posture to an auditable decision-support delivery.
Why It Matters
This project shows the kind of AI engineering I want to keep doing: using LLMs and coding agents as part of a controlled business workflow, not as a black box.
The value came from combining automation with judgment. The system could process more documents and workbooks than a manual review loop, but the outputs still needed traceability, confidence levels, and review gates so finance, operations, sales, pricing, and legal stakeholders could act safely.
Skills Demonstrated
- AI-assisted document and workbook analysis with clear human approval boundaries.
- Python/Pandas automation over messy Excel, PDF, and operational data.
- Contract and billing discrepancy analysis across multiple source systems.
- LLM/Codex subagent workflows for large-scale review, QA, and report generation.
- Product judgment: adapting from a web-app demo to the artifact format stakeholders trusted.
- Executive communication: turning raw analysis into workbooks, report language, and action-oriented delivery tasks.
Safe Public Summary
Confidential AI/data automation project for billing leakage and contract-compliance review. I combined a React/Supabase prototype, Python analysis pipelines, Codex subagents, and Excel/PDF deliverables to compare contracts, bills, payroll/support documents, and operational exports. The work produced evidence-backed review queues and executive reporting while keeping confirmed findings separate from opportunity pools and human-review cases.