Work — case studies

Each project gets
its own chapter.

Not screenshots. Stories: what was broken, what I built, what changed — with numbers where they exist and honest brackets where they don't.

PROJECT 01 — LIVE SYSTEM

EammaGloberMovers: inbox → WhatsApp

Move requests landing by email were being missed or answered hours late.

n8nGmail APIWhatsApp APIMy role: design + build + deploy

Problem

Staff checked a shared inbox manually. Peak days meant 10+ hrs/week of triage and delayed quotes — lost jobs.

Process

Mapped the inbox for a week, then built an n8n flow: Gmail trigger → parse (dates, addresses, phone) → dedupe → WhatsApp alert to the dispatcher with a one-tap reply draft. Retries + error channel included.

RESULT → Response time same-day → minutes. 0 missed alerts since launch — roughly 8–10 hrs/week of triage handed to the pipeline.

Full case study →

PROJECT 02 — AI / LLM

Support Pilot: AI ticket triage

A shared support inbox with no routing — everything read by a human first.

LLM classificationEmbeddingsPythonPostgresMy role: solo build

Problem

200+ tickets/week, no priority signal. Urgent issues waited behind password resets.

Process

Built a classifier over historical tickets (embeddings + LLM fallback), confidence-gated: high-confidence auto-routes with a drafted reply; low-confidence flags for human review. Human-in-the-loop from day one.

RESULT → 60% of tickets auto-routed; first-response time down from hours to under 30 minutes. Humans handle edge cases only.

Full case study →

PROJECT 03 — REPORTING PIPELINE

ReportFlow: 6 hrs/week → near-zero

Weekly ops reporting assembled by hand from three sources every Friday.

n8nLLM summariserSheets APIMy role: design + build

Problem

6 hours/week of copy-paste across Sheets, inbox and CRM. Errors crept in; nobody trusted the numbers.

Process

Scheduled pipeline: pull → normalise → validate → LLM executive summary with figures quoted (not invented) → deliver to Sheets + email every Monday 07:00. Validation gate blocks sends on schema mismatch.

RESULT → Manual reporting 6h → ~20min review (−94%). One source of truth, timestamped.

Full case study →

PROJECT 04 — RAG / AI AGENT

Knowledge Agent: ask your Drive

Scattered Drive files → Pinecone retrieval → an AI agent that answers and files reports to Google Docs.

n8nOpenAIGemini embeddingsPineconeDocs
$ sheets.list → loop → drive.get → pinecone.upsert … agent.answers ✓

Problem

Research questions meant 30+ minutes of open-tab archaeology across dozens of files; reports assembled by copy-paste.

Process

Ingest branch (Sheets file list → loop → Drive download → recursive split → Gemini embeddings → Pinecone) plus an answer branch: AI Agent with memory and a Pinecone Vector Store Tool, output saved to Google Docs.

RESULT → Full-corpus answers in seconds, reports auto-filed — research per report down from ~30 minutes to seconds.

Full case study →

PROJECT 05 — MEDIA PIPELINE

Launch Pipeline: form to product video

One form submission → Drive folders, background removal, AI video with status polling, Gmail delivery. Zero manual handoffs.

n8nDrive APIBG-remove APIAI video APIGmail

Problem

Every product launch repeated a dozen steps across four tools — slow, inconsistent, videos forgotten, notifications late.

Process

Form trigger → slug → folder → access → parallel upload + background removal → AI video creation → wait/poll loop until rendered → upload → update links → Gmail notification.

RESULT → Complete listing asset pack from one form fill — launch prep down from hours of manual steps to minutes.

Full case study →

Your chapter?

Your bottleneck, chapter 06.

Start a Project →