HVHoai Vi
Home / Projects / AI Reconciliation Automation

n8n + REST API + LLM

From reconciliation exceptions to an approved supplier follow-up.

A portfolio-safe automation layer built around AHF business logic. It validates sanitized exception data, asks an LLM for a structured preliminary analysis, and keeps a human approval gate before external communication.

Run the demoDownload n8n workflows
WorkflowMulti-step n8n design
IntegrationWebhook / REST / JSON
AI layerCloudflare Workers AI
SafetySanitized data + approval

Workflow architecture

A practical automation, not an AI decoration

Each stage has a clear job, a failure path and a human owner.

01Webhook intakeReceive sanitized exception JSON
02ValidateRequired fields, types and record limits
03ReconcileUse AHF exception logic and priorities
04LLM analysisSummaries, possible causes and actions
05Human approvalReview before any supplier contact
06Follow-upEmail draft, audit log and next action

Interactive demo

Run the sanitized exception workflow

The public demo never reads AHF production files. Demo mode is deterministic; live mode uses the server-side Cloudflare credentials when configured.

Loading sanitized dataset...

INPUT

Sanitized exception queue

Public-safe data
IDExceptionSKUVarianceSeverity
OUTPUT

AI-assisted review

Run the analysis to generate a preliminary review, priority actions and supplier email draft.

Human-in-the-loop

Supplier follow-up draft

Run analysis first

n8n deliverables

Three importable workflows

Each JSON file is intentionally small enough to explain in an interview and focused enough to test independently.

01

Data intake & validation

Webhook intake, schema checks, sanitization and safe rejection paths.

Download JSON
02

AI exception analysis

Calls the portfolio API and returns structured analysis for the review queue.

Download JSON
03

Approval & follow-up

Blocks outbound action until an explicit approval flag is received.

Download JSON
What I own: business requirements, reconciliation rules, exception priorities, acceptance testing and approval decisions. What AI accelerates: implementation, preliminary exception synthesis and communication drafting.

Reliability controls

Designed to fail visibly

  • Maximum 30 exceptions per API request
  • Sanitized fields and bounded text lengths
  • No API key in browser or workflow downloads
  • Deterministic fallback when live AI is unavailable
  • No supplier message without human approval

Interview proof

What this demonstrates

  • n8n workflow design and webhook intake
  • REST API and JSON payload handling
  • LLM integration through a server-side function
  • Exception handling and operational controls
  • Business-first automation ownership