AI Development Specialist · Open to remote roles

I build AI agents that do real work.

Practical AI agents and business automations with n8n, Claude & OpenAI APIs, RAG and webhooks, backed by 8+ years of web development and 40+ websites shipped.

📍 Lebanon · Remote 🕘 Available for US time-zone hours
Hala Samneh, AI Development Specialist
Hala Samneh
n8nClaude APIOpenAI APIGoHighLevelRAG / EmbeddingsSupabasePostgreSQLpgvectorTelegram BotsClickUp APIGmail APIGoogle Sheets / DriveREST APIsWebhooksPython · PandasHermesZapierMake.comClaude CodeCursor
Selected work

Automations with guardrails, not demos.

Each project starts from a business problem and ends with rules for what the AI must never do on its own.

01Built · n8n

AI Invoice Automation

Supplier invoices arrive by email as PDFs: scanned, multi-invoice, in other languages, or not invoices at all. Typing them into a ledger by hand is slow and error-prone.

What it does
A 24-node n8n workflow sends each PDF straight to Claude for structured extraction, validates the result, blocks duplicates, archives the original and writes the ledger.
Guardrails
  • Classifies every document as invoice, not invoice or uncertain before anything is written
  • Never uses a subtotal or early-payment discount as the total; never invents missing fields
  • A payment reminder keeps the original invoice number, so it is caught as a duplicate, not paid twice
  • Every file ends as Accepted, Duplicate, Exception or Failed, then one summary email per run
Tested
41 of 44 records exactly right across 4 runs of a 10-PDF test pack (a scanned invoice, a German EUR invoice, two invoices in one PDF, a payment reminder and a delivery note), scored field by field against answers written before the run. The reminder was caught as a duplicate and the delivery note rejected every time. The only miss: a vendor written with its legal suffix (“Inc.”), which showed the duplicate key needs normalised vendor names.
n8nClaude API · PDF inputGmailGoogle DriveGoogle Sheets

Demo run on the 10-PDF test pack · sample invoices, no client data

01
Gmail intakefind new emails with PDF attachments
02
Split & hash PDFsone item per file, processed one by one
03
Claude extractionvendor · number · date · total · currency
04
Validate & de-duplicatechecked against the existing ledger
05
Archive to Drive · append to Sheetssearchable file name per invoice
AcceptedDuplicateExceptionFailed
02Built · AI agent

Project Manager Bot Agent

Project briefs live in chat, tasks live in ClickUp, and blocked or overdue work is noticed too late.

What it does
A proactive agent turns a brief into ClickUp tasks with owner, due date, priority, status, dependencies and milestones, then runs scheduled checks that message owners, escalate no-response cases and produce project-health reports.
Guardrails
  • Every write starts as a dry-run preview; nothing changes in ClickUp until it is approved
  • Approval codes expire after 30 minutes
  • Only the task's assignee can mark it complete, from their own Telegram account
HermesTelegram BotClickUp APIOpenAIScheduled monitoring
Telegram chat with the PM bot: it refuses a completion from a non-assignee, previews a new ClickUp task, waits for an approval code, then confirms the task was created.

Real Telegram run: dry-run preview → approval → verified ClickUp task

03Built · n8n

AI Customer Support Email Qualifier

A makeup brand's support inbox mixes refund requests with product questions, and every email gets read and sorted by hand.

What it does
A Gmail-triggered workflow has an LLM classify each email as a refund request or a product question, extracts name, email and a one-line brief, logs it to Google Sheets and replies in the thread so the customer knows it was received.
Guardrails
A structured output parser forces the same four fields every time, the category can only be one of two values, and the model is told never to invent missing details.
Tested
36 of 36 emails classified correctly, on 2 separate runs, using 36 hand-written test emails with tricky cases (a return-policy question asked before buying, an exchange-or-refund request) and two in Lebanese Arabic. Weak spot found: names are only extracted when the customer signs the email, so the next fix is passing the sender header into the prompt.
n8nOpenAIStructured outputGmailGoogle Sheets

Test run on 36 sample emails · no emails sent

01
Gmail triggernew support email
02
LLM classificationrefund or product question
03
Structured output parsername · email · category · brief
04
Log to Google Sheetsone row per email
05
Reply in the threadpersonalised acknowledgement
04Built · RAG demo

Office Manager Bot (RAG)

Staff ask the office manager the same questions all day (leave, expenses, Wi-Fi, payday), and answers from memory drift away from the written policy.

What it does
A Telegram bot answers staff questions only from an approved knowledge base: n8n retrieves the closest policy passages from Supabase pgvector, Claude answers from those passages with the source, and every question is logged with its answer, sources and status.
Guardrails
  • If the approved documents don't contain the answer, the model must return NOT_FOUND. It never guesses
  • Unanswered questions are escalated: the employee is told, and the office manager gets the question instantly
  • Only allow-listed Telegram users can use it
Tested
30 of 30 on a 30-question test: 21 answerable questions answered correctly (one in Lebanese Arabic, answered in Arabic) and 9 out-of-scope questions escalated, with no invented answers on human review. Demo with sample policies for a fictional company, not a live deployment.
Telegram Botn8nClaude APISupabase · pgvectorOpenAI embeddings
Telegram chat with the office bot: it refuses a user who is not on the allow-list, answers the annual-leave question with its source, escalates a bonus question it cannot answer and notifies the office manager, and answers a Lebanese Arabic question in Arabic.

Real Telegram run: allow-list → sourced answer → escalation → Arabic

05Built · CRM analytics

GoHighLevel Lead & Pipeline Analytics

Marketing spend is spread across many lead sources, but nobody can see which one actually turns into closed revenue.

What it does
A GoHighLevel reporting dashboard that compares lead sources by lead volume, pipeline progression, conversion performance and closed revenue, using pipelines, opportunities and custom fields.
Note
Built on sample data during the RemotelyX AI Consultant Workshop.
GoHighLevelPipelinesOpportunitiesCustom fieldsReporting
GoHighLevel dashboard chart showing lead volume by source, from referral to LinkedIn outreach.

Lead volume by source · sample data

06Designed · client study

WhatsApp AI Assistant for a Clinic

A dermatology & laser clinic in Saudi Arabia gets more WhatsApp messages than staff can answer, and every unanswered message is a lost booking.

Design
A clinic-trained assistant that answers FAQs in Arabic and English, collects booking details, sends reminder templates and hands anything sensitive to staff with a summary.
Guardrails
  • No medical advice: every medical question goes to a doctor or staff member
  • Collects only what booking needs (Saudi PDPL treats health data as sensitive)
  • Repeats only prices and claims the clinic approved in writing
  • Hands off after two failed attempts, or whenever a patient asks for a person
WhatsApp Cloud APIn8nClaudeSupabase
01
Patient writes on WhatsApp24/7, Arabic or English
02
AI answers from the clinic knowledge basehours, services, doctors, approved prices
03
Collects booking detailsname, service, preferred time
04
Human handoff with summarymedical questions, complaints, unclear requests
05
Reminders & weekly reportapproved templates · chats, bookings, top questions
Toolkit

What I build with

AI & automation
n8nGoHighLevelClaude APIOpenAI APIAI agentsRAG / EmbeddingsVapiDeepgramElevenLabsZapierMake.com
Data & integrations
PythonPandasREST APIsWebhooksSupabase / PostgreSQLpgvectorGmail APIGoogle Sheets / Drive
Channels & tools
TelegramSlackClickUpTrelloClaude CodeOpenAI CodexCursor
Training

Programs completed

  • AI Consultant WorkshopHands-on AI consulting and business automation: n8n, GoHighLevel, APIs/webhooks, LLM workflows, Claude Code, OpenAI Codex, Make/Zapier.
  • Deployable AI Bots ProgramOpenClaw & Hermes agents, RAG, Telegram/Slack, ClickUp integrations, scheduled monitoring, escalation logic and guardrails, with a practical build assessment.
  • AI Consultant InternshipTurned business requirements into automation and agent solutions; tested integrations and documented workflows.
Contact

Have a process that should run itself?

I'm open to remote AI automation and AI agent roles, available for US time-zone hours.