Investigation & Reconciliation Agents
Traced a $48k reconciliation gap to its source, and proved a separate $5,989.70 exposure was real, growing about 10x, and scoped correctly.
The problem
Two companion systems, both built to replace "open six spreadsheets and eyeball it" with something repeatable: an on-demand investigation agent for support and finance escalations (plain-English description in, structured root-cause report out), and a deterministic reconciliation pipeline for periodic settlement and loyalty-balance validation. The value is not the automation for its own sake. It is that both systems have found real, dollar-denominated problems and, just as importantly, cleared real problems that turned out not to be problems, with evidence either way.
What it found
- Traced a merchant’s claimed $48k "overfunding" complaint to its actual source: 38 specific bank deposits carrying a routing signature that did not match any Wedge payout record. Proved our own ledger was correct to the dollar ($151,337 recorded vs. $152,347 the merchant’s bank showed, reconciling exactly within normal ACH timing) and identified the real cause as a partner-side routing mismatch.
- Sized a slow-burn ACH exposure precisely: $5,989.70 lost to a specific failure mode across 113 events all-time. The more important finding was that the monthly run rate had grown roughly 10x (from about $90/month to about $963/month), reframing a "$6k lifetime, do not bother" story into an "about $11.6k/yr and climbing" story worth fixing.
- Reconciled loyalty cashback balances exactly across roughly 1,023 customer accounts for a two-week period, to $0.00 variance.
Stack
Claude Code Workflow scripts (multi-phase: ingest, reconcile, report), containerized database access with pre-mounted credentials, Python sanitization and reconciliation scripts with deterministic PII hashing.