srikomma-netizen

Payment Integrity Analyst — Data skill for Claude Code

Data community

Natural-language payment-integrity analyst agent on Claude: text-to-SQL over a healthcare claims and investigations mart with a deterministic SQL guard, PHI column controls, result-level evals, SIU-le.

How to install Payment Integrity Analyst

This entry records only its repository, not the path inside it, so there is no exact command to give. Open srikomma-netizen/finance-analyst-agent and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Payment Integrity Analyst does

Natural-language payment-integrity analyst agent on Claude: text-to-SQL over a healthcare claims and investigations mart with a deterministic SQL guard, PHI column controls, result-level evals, SIU-lead approval interrupts, and agentic policy-document navigation.

Alternatives in Data

  • SQL Queries — Generate SQL queries from natural language across major dialects 7.8k ★
  • Token Dashboard — See where Claude Code is burning tokens - turn raw JSONL transcripts into local cost analytics, hotspot views 669 ★
  • Design Gdpr — Design GDPR/CCPA-compliant consent flows — cookie banners, consent hierarchy, privacy controls, and data delet 309 ★

README

Payment Integrity Analyst Agent

A natural-language analyst for a healthcare payment-integrity mart: claims, payments, providers, vendors, risk flags, and investigation cases. Built the way a regulated team would ship it: the model decides *what* to query, deterministic code decides *whether it may run*, the database decides *what the numbers are*, and the model is only allowed to narrate numbers that exist. PHI columns cannot be queried at all.

Everything in this repo is synthetic. No client code, no client data, no real PHI.

question ─▶ understand ─▶ guard ─▶ (SIU-lead approval?) ─▶ execute ─▶ compose ─▶ verify ─▶ answer
                │            │
                │            └─ rejected ─▶ back to understand with the reason (max 3 attempts)
                ├─ ambiguous ─▶ ask one clarifying question
                └─ not in catalog / not permitted ─▶ refuse

Typical questions: how many claims each risk rule flagged last month, the false-positive rate of closed cases by rule, duplicate-payment overpayment totals, investigation cycle time by outcome, which vendors had unverified bank-change flags.

What is in here

Path What it shows
analyst/warehouse.py Synthetic SQLite payment-integrity mart (members, providers, vendors, claims, payments, risk_flags, cases, restricted investigator_notes), read-only executor with timeout and row cap
analyst/catalog.py Semantic layer: table and column docs, metric definitions (flag rate, false-positive rate, duplicate overpayment, days to close, recoveries), PHI columns marked restricted and hidden from the model
analyst/guard.py Deterministic SQL guard on a real parser: SELECT-only, table allow-list, role permissions, restricted PHI columns anywhere in the query, dangerous functions, LIMIT enforcement
analyst/llm.py Claude via the official SDK with structured outputs, plus a deterministic fake behind the same interface
analyst/workflow.py ~100-line state-gra