Build an ARR reconciliation workflow for a finance/revenue-operations team. Its job is to compare annual recurring revenue between the CRM (Salesforce — or your CRM) and the ERP/accounting system (NetSuite — or your billing/ERP), match accounts by their shared account ID, surface the variances that matter, and give reviewers a place to categorize each one, add notes, and mark a month as reviewed. Trigger it two ways: a manual/API run that refreshes the data on demand, and interactive HTTP pages that people open in a browser. Keep the routes private to the space by default. At a high level, build these pieces: A fetch-and-reconcile pipeline: pull active accounts and their ARR from the CRM, pull the corresponding revenue figures from the ERP (from line-level revenue elements, as-of the end of the reconciliation month), join them by account ID, compute the value and percentage variance per account, and persist the results per monthly period to durable storage (a SQLite database on a named volume works well). Reads from source systems are read-only. A main dashboard (web page) listing every reconciled account with both ARR figures, the variance, an assigned category ("bucket"), reviewer notes, and an expected-resolution month. Support sorting, filtering, and search. A row should only be flagged for review when the variance crosses a materiality threshold (both a dollar floor and a percentage floor) — keep the threshold easy to change. Also read the account type from the CRM and, when an account is marked churned there, auto-tag it and its bucket as churn. Let reviewers lock a month once it's reviewed so it stops being overwritten by refreshes. A customer detail page reachable from the dashboard, showing one account's revenue elements from the ERP (including the source sales-order reference), its comment thread, and a per-account view of the variance. An AI variance-analysis step that inspects the ERP revenue elements for accounts with a material variance and proposes a likely cause and suggested fix, shown on both the dashboard and the detail page. Use a capable LLM (Claude — or your preferred model provider). Small supporting API/handler steps for reading rows, updating a row, posting comments, and locking a period. Store reviewer edits and categories durably so they survive refreshes, and make the pipeline safe to re-run. Tools used: Salesforce API, NetSuite SuiteQL (REST), Claude (Anthropic) API, 3B SQLite named-volume storage