Worker Builder

From plain words to a tested Worker

Build a Worker by describing the work, not by writing code. Each instruction becomes part of its configuration and a step in its graph, ready to test in eVal.

Get in touch

Step 01 · Build

Worker BuilderNew WorkerSubmission triage DraftReady to test

Let’s build a Worker

Describe the work in plain words.

Connected Broker submissions Underwriting guidelines Broker master

Build a Worker that triages new marine cargo submissions.

Worker Builder

Created Submission triage. It starts on each new submission in Broker submissions and reads the pack.

Add a step that checks the requested limit against the Authority limit

Worker Builder

Found it in Guidelines v12, §4.1: €3.0m for marine cargo.

Add a branch: over authority goes to a senior underwriter, citing the guideline.

Worker Builder

Added. Everything else goes to the underwriting queue. Ready to test.

Search your data

Broker submissions

Submission pack

Underwriting guidelines v12

Broker master

Authority limitGuidelines v12, §4.1

Authority matrixUnderwriting

Authorised brokersBroker master

Describe what the Worker should do Type @ to search your data

Build a Worker that triages new marine cargo submissions.

Add a step that checks the requested limit against the @authAuthority limit

Add a branch: over authority goes to a senior underwriter, citing the guideline.

New WorkerSubmission triage submission-triage.yaml · 7 lines11 lines16 lines

Trigger

None yet New submission in Broker submissions Connected

Inputs

None yet

Requested limitConfirmed
Submission pack · schedule.requested_limit
SUB-2291 · Northgate Trading · €4.2m

Authority limitConfirmed
Guidelines v12 · §4.1
€3.0m · marine cargo

Rules

None yet

1Over the authority limit in Guidelines v12, §4.1, refer it to a senior underwriter.

2Anything else goes to the underwriting queue.

PreviewLLMIfCode

New submission LLMRead the pack CodeCheck the limit Over authority? yes no Senior underwriter Underwriting queue Cites §4.1

Prompts, not clicks

Tell Worker Builder what the Worker should do, and it sets up the trigger, the steps and the rules as you go. It builds in the context of the data and schemas it has access to, so every input points to a real source.

Step 02 · Edit

MellonWorkers/Submission triage/RulesVersion 4.2 · liveVersion 4.3 · draft

Worker BuilderSubmission triageVersion 4.2 · liveVersion 4.3 · draft

Rules

Plain English. The first rule that matches decides.

SaveSaved

1Over €15m€10m, send it to the head of marine.Edited
2Over the authority limit in Guidelines v12, §4.1, refer it to a senior underwriter.
3Anything else goes to the underwriting queue.

Version 4.3, a draft

Rule 1: €15m → €10m
Configuration: 1 line changed
Graph: Same shape

Goes live once it passes eVal

Business logic configurability at your fingertips

The people who own the process change the Worker themselves. Rewrite a rule in plain words and the draft updates at once: one line changes, and the rest of the Worker stays as it was.

Step 03 · Test

MellonWorkers/Submission triage/TestsVersion 4.3

Worker BuilderSubmission triageVersion 4.3

Simulations

48 from the Worker's rules, 12 of them edge cases

Running1 failed · 47 of 48 pass48 of 48 pass

€4.2m, over authoritySenior underwriter
Exactly €3.0mEdge caseUnderwriting queue
£2.8m, in sterlingEdge caseWent to the queueSenior underwriter
Schedule €3.2m, email €2.9mEdge caseSenior underwriter

One edge case failedOne edge case fixed

£2.8m was read as €2.8m and went to the queue. Proposed fix: convert the limit to euros in code, before the check.

Apply fixApplied

Every path tested, edge cases included

Worker Builder generates simulations from the Worker's own rules, the awkward cases at the edges among them, and runs every one. When a case fails, it shows why and proposes a fix for you to approve, so each round leaves the Worker harder to break.

Step 04 · eVal

MellonEval/Evaluations/Submission triagePass mark 80

eValSubmission triagePass mark 80

Submission triage

Golden set v3, settled by three underwriters · 4,200 runs a version

VersionScore against the pass markScoreGate

v4.0Baseline71Not released

v4.1Cheaper model54Rejected

v4.2Knowledge layer context84Contained release

v4.3Two steps moved into code91Running 4,200 casesReleased

Illustrative run. Scores are out of 100; the pass mark is 80.

Every version scored before it reaches your desk

Each version changes one thing and runs against cases your own experts have already settled. If it misses the pass mark, it doesn't ship.

Four versions, one exam

Versions 4.0 to 4.3 each took 4,200 runs on the same settled cases. A cheaper model scored 54 and was rejected; moving two steps into code took the Worker to 91.

See Mellon across your organisation

Tell us about your organisation, and the right person on our team will be in touch.

Get in touch

Security and control, built in

Mellon runs in the cloud, under keys you hold. Its Workers read your systems and write to them only with your permission. Every run is sealed in a record you can access in full, at any time.