Human-in-the-Loop2025 to 2026

Claims Ops Platform

An operations platform that reconciles campaign claims with banking and insurance partners, replacing a manual, email-driven process.

Watch the explainer: Claims Ops Platform, 75 seconds

Watch the 75-second narrated explainer. English captions on.Read the transcript

TL;DR

Gated pipelines in both directions, transform rules that can be generated with the Claude Agent SDK, and a jobs layer built after a real 504.

Role
Platform Founder & Integration Owner
Timeline
2025 to 2026 · Ongoing
Stack
  • TypeScript
  • Fastify
  • Postgres 16
  • React
  • n8n
  • Claude Agent SDK

01

Context

Full automation was the wrong goal. Partner data is messy, formats drift and a wrong export has financial consequences, so the system automates the mechanical 90 percent while making human judgment explicit, auditable and impossible to skip by accident.

02

What I built

  1. Two pipelines: outbound claims with an explicit mark-ready gate before any CSV can be exported, and inbound partner reports with per-row mismatch review and operator override.
  2. Transform rules three ways: passthrough, hand-written, or AI-generated with the Claude Agent SDK.
  3. An async job layer born from a real 504: a DB-backed jobs table with a typed runner owning all state transitions, errors persisted to the row so they can never crash the process.
  4. A dual-pull adapter mapping a parallel v2 campaign system into the v1 record shape, so every existing gate applied unchanged, with non-fatal degradation if v2 fails.
  5. Governance of the company's n8n fleet: an instance manager with cross-instance workflow transfer, sub-workflow ID remapping and ownership-aware migrations.

03

Engineering judgment

Integration boundaries were engineered against observed behavior rather than documented contracts: upstream quirks are verified empirically, then encoded as constants, comments and tests so the finding outlives any individual engineer. The same philosophy governs every gate in the pipeline: warn without blocking on deadline risk, and report rather than silently drop excluded data.

04

Outcome

The platform became the ops team's daily system. Two teammates have since added hundreds of commits on the skeleton, while I continue to own the infrastructure and the upstream integration contracts.

05

Transcript

Read the 75-second explainer as text

Full automation was the wrong goal. Partner data is messy, formats drift, and a wrong export costs real money. The Claims Ops Platform reconciles claims with banking and insurance partners. It automates the mechanical 90 percent, and makes human judgment impossible to skip by accident.

Outbound, claims are transformed into each partner's format, but no CSV exports until an operator marks it ready. Inbound, partner reports are checked row by row, and each mismatch waits for operator review. Transform rules come three ways: passthrough, hand-written, or generated with the Claude Agent SDK.

Then a real request hit a 504. The fix: a jobs table in the database, a typed runner that owns every state change, and errors saved to the row, so they never crash the process. Built in TypeScript on Fastify, Postgres and React, with n8n and the Claude Agent SDK. It became the ops team's daily system, and 2 teammates have since built on its skeleton.

Automate the mechanical. Make judgment explicit.

Work with me

Have a system like this that needs to reach production?

Book an intro call to talk it through, or email me a few lines about what you are building.