AI config authoring
When I was selling, the slowest part of getting a customer live was waiting on configs, the files that tell Sensible where each field sits on a document. Every new format needed a solutions engineer to write one by hand, and I wanted to see whether AI could write the first draft.
Learn from what our engineers already shipped
Instead of teaching a model our documentation, I gathered a few hundred production configs our engineers had written for customers, along with their sample documents, and made them searchable. Each one has judgment calls the docs never spell out, like which method holds up on a scanned page.
golden_samples/
Financial Services/ bank statements, balance sheets, VOEs
Insurance/ ACORD 25/28, loss runs, declarations
Healthcare/ EOBs, CMS 1500
Tax Forms/ 1040s, W-2s
Human Resources/ pay stubs
Real Estate/ closing disclosures, title
Utilities & Invoices/
golden_samples_index.json searchable index of every config
reasoning_patterns/ patterns pulled out of the configs
.claude/agents/ the four agentsSplit the job so it can't cut corners
When one agent did everything, it would copy a sample config and call it done. So I split it into four: one groups the documents, one studies the layout and writes a plan, one writes the config, and one tests it against the API and sends failures back to whoever caused them. The one writing can't test its own work.
- ScoperGroups the PDFs
- StrategistReads the layout
- CoderWrites SenseML
- QATests against the API
Each agent is a short file of rules
Each agent is a markdown file with its job, the tools it's allowed and the tools it isn't. Taking tools away was the part that made it work. The coder can read and write files but can't run anything, so it has to hand off to QA.
---
name: senseml-coder
description: Write or edit a SenseML extraction config from a strategy memo.
tools: Read, Write, Edit, Grep, Glob
disallowedTools: Bash, Task
skills: senseml-methods, senseml-patterns
---
## Your Workflow
1. **Receive** the Strategy Memo from the Orchestrator
2. **Read** the reference config identified by the Strategizer
3. **Write** the SenseML JSON configEngineering took it from there
It matched a new-hire test config field for field, then built a full set of production configs for a real customer out of a pile of mixed insurance PDFs. That's when it became a company project. Our engineering lead took the main lesson, that a lot of this is really search and simulation, and built a toolkit that tries candidate configs on a local copy of the engine. It's being built into the product now, and our solutions engineers already use it on customer builds.
Being integrated into the product, built on the same ideas.