Automation Lab

Production workflows and AI agents built in n8n. Every one of these replaced something a person used to do by hand.

Telegram AI agent · client project2025 — present

Teacher Assistant AI Bot

A natural-language Arabic assistant delivered to a private tutor in Saudi Arabia. The tutor manages students, schedules, cancellations, and dues entirely through chat — the AI agent calls custom CRUD tools under strict business rules rather than guessing.

Custom tool-calling design with strict business rules
Usage metering and free-trial gating
Session memory across conversations
Automated lesson reminders scheduled in Saudi time

Outcome: A tutor runs an entire practice from a chat window, with no app to learn and no spreadsheet to maintain.

Production document AI pipeline2026

Smart Scanner Extraction Pipeline

The production pipeline behind Smart Scanner. Every invoice — photo or PDF — is verified as a real invoice, parsed with GPT-4.1 Vision, validated, deduplicated, and persisted. It reads Arabic, English, and mixed documents, and is tuned for handwritten Arabic invoices and ZATCA e-invoice fields.

Rejects any file that is not a genuine invoice, before anything is saved
Extracts 70+ fields plus line items: vendor, buyer, totals, VAT, dates
Classifies every line item into per-company expense categories at extraction time
Refuses politely and explains why when required data is missing — accuracy before speed

Outcome: Manual invoice entry disappears entirely; a photo becomes structured accounting data in about ten seconds.

Scheduled agent with human-in-the-loop2026

AI Job-Match Pipeline

A scheduled pipeline that runs five daily search queries across LinkedIn, Indeed, and Glassdoor, scores each role's fit from one to ten with a structured-JSON LLM evaluator, and delivers only the high scorers.

Structured-JSON evaluator scoring fit on a 1–10 scale
Deduplication against n8n Data Tables so nothing is seen twice
Telegram delivery of high-scoring matches only
Human-in-the-loop approval before any application is sent
Semi-automated applications with tailored AI-generated cover letters via Gmail

Outcome: Daily job hunting collapses into reviewing a short, pre-scored shortlist.

Scraping and enrichment pipeline2025

Smart Lead Finder

Automated B2B lead generation across the Saudi and Jordanian markets: multi-query place search, website scraping to extract emails and WhatsApp numbers, prioritisation scoring, and duplicate-safe appends into segmented sheets.

Multi-query place search across two markets
Extracts emails and WhatsApp numbers from company websites
Prioritisation scoring so the best prospects surface first
Duplicate-safe appends into segmented sheets

Outcome: Manual prospect research became a hands-off scheduled pipeline producing scored, deduplicated lead lists.

Video-to-data webhook service2026

Lesson Quality Evaluator

A webhook service that transcribes lesson videos with speaker separation and returns structured JSON scoring teacher quality and student engagement.

Speaker diarization separates teacher from student
Structured outputs guarantee a parseable result every time
Returns a score, not a transcript — the caller gets data, not homework

Outcome: Lesson review becomes measurable instead of anecdotal.

Data quality pipeline · proof of concept2026

CRM Lead Cleaner

Incoming CRM leads arrive scattered — duplicates, malformed emails, missing fields — and cost sales teams hours of manual triage. This pipeline cleans them automatically and reports on what it did.

Removes duplicates and validates email addresses
Flags missing fields rather than silently dropping records
Splits results into Clean and Rejected, with a reason on every rejection
Produces a quality summary after each run

Outcome: Designed to plug straight into a CRM, turning hours of manual triage into a summary line.

A case study in cost: one branch, three quarters saved

$0.02 per PDF, vision model
$0.005 per PDF, text model
75% cut per document

Webhook — file arrives PDF with extractable text?

Yes Text model reads it $0.005
No Vision model sees it $0.02

Extraction came back empty? The PDF was only a scan — send it back to the vision branch.

1,000 receipts a month, half PDF, half photos ~~$20~~ $12.50

The mistake in my first workflow was sending every file that arrived at the webhook straight to a vision model. It worked. It also cost roughly three times more than it needed to.
The reason is simple: a PDF usually already contains readable text. A vision model does not need to look at the page when a text model can read it — and reading costs a fraction of seeing.
The fix was a single branching decision placed just after the webhook: if the file is a PDF with extractable text, send the text to a text model; otherwise send the image to the vision model. The per-document cost dropped by about three quarters.
One detail matters more than the branch itself: always keep a fallback. If text extraction returns empty — because the PDF is really just a scanned image — the file has to go back to the vision branch. Without that, the cheap path silently produces nothing.