← All projects

LIVE DEMO

Quote Lifecycle Desk

n8n follows quotes across scheduled emails, replies, reviews and failure paths without duplicating a milestone

STACKn8n 2.31.6 / PostgreSQL 16.14 / Google Sheets API / Gmail API / OpenRouter structured output / Next.js / Docker

CONTEXTSelf-directed spec project

DATE2026-07

Open recruiter demo ↗ quote-lifecycle-desk-demo.aaron.cv

EVIDENCE, NOT CLAIMS

The lifecycle risks a reviewer can trigger

The public console states the expected policy result before every run, then shows structured classification, confidence, guardrail, Sydney business-hours routing, idempotency, retries, audit ID, execution trace and the stored synthetic state. The public parser is deterministic and makes no model call. The 64-case and 8-case figures both come from the hash-locked release evidence measured on 1 August 2026; the accepted delivery run later that day wrote its own live results to the current-run path and left the published figures untouched.

64/64 live model responses passed the structured schema Versioned 64-case acceptance matrix

64/64 guarded dispositions matched the acceptance oracle Deterministic validation after structured output

22/22 ambiguous and malformed cases routed to review No sales outcome accepted for these cases

56/64 raw provider classifications matched before guardrails Reported separately from guarded system quality

20 + 20 concurrent scheduler and dispatch workers in the shared-milestone test One action, outbox row, effect and lifecycle event

8/8 behaviour classes passed in the selected release run 1 August 2026 at 21:57 UTC; 2,739 tokens, no provider errors

11/11 acceptance gates passed in the accepted delivery run 1 August 2026 at 23:05 UTC; two gates compare the installed instance with the build

Not measured production accuracy, throughput and business impact Synthetic reference traffic only

A milestone is not sent twice

A unique database key, transactional lease and stored action receipt suppress repeated scheduler and webhook effects.

VERIFY: Run Duplicate milestone and inspect the two calls in the raw response

A reply resolves to one quote or none

Automatic processing requires the stored Gmail thread and visible quote reference to agree on one record.

VERIFY: The source suite covers exact, missing, wrong, multiple and conflicting references

A promised date moves the next action

A valid deferred decision schedules one check at 09:00 Australia/Sydney and cancels the previous pending check.

VERIFY: The acceptance suite verifies the deterministic deferred_check state

Uncertainty cannot become a sales outcome

Low confidence, ambiguity, malformed schema and contradictory evidence leave the quote open and create review.

VERIFY: Run Low-confidence reply and inspect state.reviews

The report is replayable

Totals and loss reasons come from stored events and an event watermark; the same input produces byte-identical JSON and hash.

VERIFY: Download the measured evidence summary

Failures are bounded

Retryable mail and carrier errors stop after three total attempts and create one dead letter per source key.

VERIFY: Run Provider failure and inspect retry count, fallback and stored dead letter

The installed instance must match the build

Delivery compares every installed n8n workflow node for node with the versioned builders and requires the four named credentials to resolve. A drifted workflow, or a credential a test run left behind, fails the release check instead of passing behind green tests.

VERIFY: Compare delivery verification and independent review in the measured evidence summary

Published claims stay tied to one accepted release

The accepted evidence locks four source hashes and all six must-prove mappings. Both live suites now move only through an explicit snapshot selection, so a routine provider run cannot age a published figure, and the replaced matrix stays committed next to it. A separate export gate compares the sanitized 11-workflow set and its 97 nodes with the versioned builders.

VERIFY: Download the measured evidence summary from the proof console

TRY THESE CASES

Routine acceptanceWon, normal priority, confidence above 0.85 and a stored lifecycle update

Urgent exceptionHuman handoff, commercial escalation and no stored sales outcome

Low-confidence replyConfidence below 0.85, quote stays open and one review record

After-hours exceptionSydney business hours false and next-business-day review routing

Duplicate milestoneFirst event handled, second suppressed and one audit reference

Provider failureThree attempts, one dead letter and human review fallback

Custom synthetic replyThe same schema, 0.85 confidence gate, routing rules and audit check apply

RUNTIME PATH

Public proof console validates a synthetic reply and calls one internal n8n webhook
n8n returns structured lifecycle category, priority, confidence and Sydney business-hours state
A deterministic gate chooses state change, human review or dead-letter fallback
The maintained public workflow executes versioned PostgreSQL transaction functions
PostgreSQL stores quote state, idempotency receipt, audit events, reviews and dead letters
Private acceptance profiles test Google Sheets, Gmail threading and model output without exposing credentials

DATA BOUNDARY

Synthetic data only in the public demo. The browser can reach the proof console, not n8n, PostgreSQL or provider credentials.

The public custom-input path uses a deterministic sandbox parser and does not call a model provider.
Live acceptance used private Gmail, Sheets and model profiles; those credentials are absent from the demo stack.
Provider processing region and retention were not configured or measured.
Workflow executions omit success and error payload storage; synthetic state resets nightly.
No client data, client tenant, revenue result, conversion result or saved-time claim is part of this project.

Long-running quotes create a coordination problem: the follow-up date lives in one place, the email thread in another and the eventual decision often arrives as unstructured text. Quote Lifecycle Desk treats the lifecycle as stored events instead of a set of editable status cells.

A real Google Sheets acceptance profile reads and writes synthetic quote rows. n8n schedules day-10 follow-up and day-21 escalation, then Gmail replies must match both the stored thread and the visible quote reference before a record can change. PostgreSQL owns the state machine, unique milestone claims, outbox, review queue, audit trail and dead letters.

Reply classification uses a strict versioned schema and a deterministic gate. Confidence below 0.85, malformed output, contradictory evidence or an invalid decision date leaves the quote unchanged and creates human review. Weekly reporting is replayed from stored events and an event watermark, so the same event set produces the same content hash.

This is a self-directed spec project with synthetic data, not client work. The public console runs the real n8n sandbox workflow and PostgreSQL transactions. Live Gmail, Sheets and model credentials stay outside the public stack; their results are reported only from the versioned acceptance evidence.

What it does

Day-10 follow-up and day-21 escalation with unique milestone claims under concurrent scheduler runs
Reply matching requires one Gmail thread and the same visible quote reference before any state change
Won, deferred and four loss outcomes behind schema, evidence, confidence and date guardrails
Ambiguous or malformed replies create review records without changing the sales outcome
Weekly summary reproduced from stored events, event watermark and SHA-256 content hash
Mail and carrier failures stop after three attempts and end in an auditable dead letter