Built a 5-agent AI system that cut 20-30 hours of weekly manual work to under 4 hours

What I Built

1. Multi-Agent Email Intelligence Pipeline

I designed and deployed a 5-agent AI system in n8n that processes inbound customer emails end-to-end — no human intervention until final approval. The pipeline works like this: one agent scores sentiment, another retrieves answers from a Gemini-powered knowledge base, a third drafts a response in my exact brand voice (analyzed from my writing samples via a custom Echo Workflow), a fourth runs quality assurance, and a fifth posts the draft to Slack for one-click approve/edit/reject. A separate approval handler workflow watches for the Slack reaction and either sends the email via Gmail, opens an editor, or discards it. Five AI agents, two workflows, six API integrations — fully automated with human-in-the-loop governance.

2. LP Analyst Platform — Fraud Detection at Scale

I built three standalone HTML5 analytical tools that replaced 10-15 hours per week of manual order review with 2-3 hours of targeted investigation. These tools ingest Excel exports from two different POS systems (Tiger Detail and KEYS Orders), join and deduplicate the data client-side, then run multi-factor risk scoring models I designed from scratch. The delivery risk tool alone evaluates 7 high-risk indicators per order — low delivery times, self-service patterns, cash payment anomalies, quick-update manipulation — with tiered volume multipliers that properly weight systematic fraud. I added a second tab that detects management manipulation by analyzing order timestamps against driver shift checkout/clockout times, and a third that surfaces un-dispatch fraud through blank driver ID analysis. These tools catch 3x more fraud patterns than manual review, including an entirely new category of insider manipulation that was previously invisible.

3. Chargeback Report Automation

A 17-node n8n workflow that replaced a 3-hour weekly manual process with zero-touch automation. Every Tuesday, it pulls chargeback Excel files, fetches store-to-market mappings from Airtable (the single source of truth), normalizes and deduplicates the data, splits by market (DC vs NY), generates formatted PDF summary tables, and emails them to the correct recipient lists. The entire pipeline — from raw Excel to delivered PDF reports — runs without a single click.

4. Cloud-Resilient Data Ingestion Pipeline

I architected an ETL pipeline that pulls KEYS Orders and Tiger Detail files from OneDrive via the Microsoft Graph API, parses the Excel data, normalizes field names across two different data schemas, deduplicates records, enriches them with metadata, and outputs clean structured data for downstream risk scoring. This solved a critical business continuity problem — the old process depended on a local machine being online. I discovered and fixed several non-obvious data engineering challenges along the way, including Excel's unnamed column behavior (EMPTY fields), business-date vs. calendar-date handling for 24/7 operations, and deduplication ordering bugs that were inflating fraud counts.

5. Brand Voice System

I built a workflow that analyzes writing samples through Gemini AI and outputs a structured XML brand voice profile — capturing sentence patterns, vocabulary preferences, tone signatures, and communication archetypes. This profile powers the email drafting agent so every AI-generated response sounds authentically like me, not like a chatbot.

6. Knowledge Base Infrastructure

I deployed a complete Gemini File Search corpus — menu data, FAQs, location info, policies, brand voice — so AI agents can retrieve accurate, grounded answers instead of hallucinating. When Google deprecated the API endpoint I was using, I engineered a workaround using the Files API directly and built Python verification scripts to confirm upload integrity.

7. LP Disposition Collection System

I designed a JotForm-to-n8n-to-Airtable workflow that replaces the phone tag process for collecting case dispositions. Target impact: 5-10 hours/week of back-and-forth calls reduced to 30 minutes, with average turnaround dropping from 5 days to 2.

The Numbers

MetricBeforeAfterWeekly LP analyst review10-15 hours2-3 hoursChargeback report generation3 hours/week0 (fully automated)Disposition collection5-10 hours/week30 minutesFraud pattern detectionBaseline3x improvementASE manipulation detection0% (invisible)100% visibilityEmail response turnaround2-5 daysSame dayReport delivery reliabilityVariable100% on-timeTotal manual hours saved~20-30 hrs/week~2.5-3.5 hrs/weekThe Technology Stack I Command

- Workflow Orchestration: n8n (50+ production workflows, multi-agent pipelines, webhook/schedule triggers, conditional branching, error handling)

- APIs Integrated: Microsoft Graph (OneDrive), Google Gemini (RAG + File Search), Airtable, OpenRouter (LLM routing), Slack, Gmail OAuth2

- Data Engineering: Python (pandas, python-docx) and JavaScript (SheetJS, Lodash) for ETL, normalization, deduplication, percentile calculations, and report generation

- Frontend Tools: Standalone HTML5 applications with client-side Excel parsing, multi-tab interfaces, sortable/filterable tables

- AI/ML: Multi-agent orchestration, brand voice analysis, sentiment scoring, risk scoring models, RAG-powered retrieval

- Infrastructure: Airtable as operational database, Gemini as knowledge store, Slack as approval interface, cloud-first architecture

What Makes This an Architecture, Not Just Automation

Individual automations save time. What I built is an architecture — a composable system where each piece connects to and strengthens the others:

- Airtable serves as the single source of truth across workflows (store mappings, case data, client configurations)

- The brand voice profile feeds into the email agent, which feeds into the Slack approval loop

- The data ingestion pipeline feeds the risk scoring engine, which feeds the analyst tools, which feed investigation workflows

- The knowledge base powers customer-facing AI responses AND internal operational queries

- Everything is designed to scale from 1 client to 12 with configuration changes, not rebuilds

The platform architecture supports projected recurring revenue of $1,800-3,600/month per client at 89-97% gross margins — because once the system is built, onboarding a new client is configuration, not development.

The Approach

Every system I built followed the same methodology:

1. Identify the pain — Find the manual process burning hours

2. Map the data flow — Understand inputs, transformations, and outputs

3. Design the automation — Architect the workflow before building it

4. Build incrementally — Test each node/component before connecting them

5. Handle the edge cases — Business dates, unnamed columns, duplicate records, deprecated APIs

6. Document everything — 146 markdown files ensuring anyone can maintain and extend the work

7. Validate with real data — Every system tested against production datasets

I didn't just learn AI tools. I learned to think in systems — to see a manual process and immediately envision the workflow, the data transformations, the API calls, the error handling, and the human touchpoints that turn it into something that runs itself.

Built November 2025 – February 2026 as part of the MindValley AI Mastery program and deployed for Progressive Loss Prevention Concepts.