Arabic AI chatbot and automation for customer support and repetitive work
- Typical timeline
- 3–6 weeks
- Deliverables
- 8
- Updated
- 2026-10-05
We build an Arabic AI chatbot and automation for repetitive tasks: an assistant that answers your customers in Arabic from your own business information, and tools that do in bulk what someone on your team now does by hand, in a browser or in Excel files. Mahmoud Hamdy built both of the examples below: IAI, a support assistant for a Saudi 3D-printing store on Salla, and Salamah Checker, a bulk permit checker for Esnad Al-Mudun on the Saudi Civil Defense Salamah portal.
Inside IAI, an Arabic assistant for a Salla store
The store's customers ask in Arabic about products and about their orders. The brief was that answers had to come from the store's own information and current catalog, not from a general chatbot making things up.
Mahmoud built a Node.js and TypeScript service that takes questions from a chat widget on the store and from a Telegram bot. It calls the OpenAI and Anthropic APIs and looks things up in a knowledge base with semantic search, which matches on meaning rather than exact words: a question like "when will my order arrive?" can find its answer in an entry written under "shipping times". Frequent answers are cached so the model is not called again for a question it has already answered. The whole thing is deployed with Docker.
Before any reply is written, a question goes through two-stage intent classification. A fast keyword filter narrows the options, then the model scores the question against known intents: an order inquiry, a product search, a common question, shipping, returns, or a request to talk to a person.
What the assistant knows, and when it hands over to a person
In IAI a confidence threshold decides whether the assistant answers on its own or escalates the chat to a person. The store owner sets the escalation rules and the reply templates from the dashboard, with no code changes.
The assistant does not guarantee a correct answer every time; language models can be wrong. That is why the decision is not left to the model alone: IAI relies on the confidence threshold, the hand-over to a person, and a knowledge base the store keeps up to date. We publish no usage numbers for IAI, because we do not have documented figures we can show.
Connecting to Salla: what we have and have not built
IAI reads products, prices, stock and order status through the Salla API, by polling and through the webhooks Salla sends when something changes. That is how it can answer "is this in stock?" from today's data rather than a stale copy.
To be clear about scope: we have built one assistant connected to one Salla store. We have not built Salla themes or apps for the Salla app store. If that is what you need, it is not work we have done before.
Browser automation at scale: 30 sessions in Salamah Checker
Esnad Al-Mudun was checking permit statuses on the Salamah portal one at a time. In a two-month freelance engagement, Mahmoud built a Windows desktop tool in Python and Playwright that takes permits by number, by range or from an Excel file and spreads them across up to 30 parallel browser sessions.
Long runs get interrupted, and government portals sometimes return temporary errors. The tool saves checkpoints so a run can pause and pick up where it stopped, and it retries failed checks with growing delays instead of hammering the portal. Results go into SQLite, a single file on the user's machine with no server to run, and export to Excel.
It shipped as an installer and as a portable build that runs without installing, covered by 179 automated tests.
Offline data processing: the Python pipeline behind Ihkam
Not every automation needs a language model. For Ihkam, a platform for Quran memorizers at ihkamquran.com, Mahmoud wrote a Python pipeline that runs ahead of time rather than on each request. It pins the Tanzil Uthmani text with a hash per ayah, finds repeated word sequences across all 6,236 ayat, groups them, classifies each difference (a particle, an addition, an omission, a substitution, word order or a verse ending) and generates 43,201 practice questions.
No language model is involved: the Quran text must never be generated or altered by a machine. The web app only reads what the pipeline produced.
When an Arabic AI chatbot and automation make sense, and how long they take
A chatbot fits a store that gets the same questions about products and order status again and again, as with IAI. Automation fits manual work repeated across many records, like the one-at-a-time permit checks that Salamah Checker replaced.
We plan an assistant or an automation tool as an add-on to a business that already runs: 3 to 6 weeks, or 6 to 10 weeks as part of a full first release with a mobile app and a dashboard. Salamah Checker took two months, tests and both builds included.
We are two engineers in Menoufia, Egypt, working remotely for clients in Egypt and the Gulf. Mahmoud builds these tools and is the person you talk to about them.
- Typical timeline
- 3–6 weeks
- Tools we use
- 12
- Projects
- 2
What you get
- An assistant service in Node.js and TypeScript on the OpenAI or Anthropic APIs
- Chat channels: a widget on your site or store, and a Telegram bot
- A semantic-search knowledge base built from your business's information, with a cache for frequent answers
- A connection to your live data, such as products and orders in a Salla store
- A Next.js dashboard for conversation logs, reply templates and the rules for handing a chat to a person
- A Python and Playwright automation tool with pause and resume, retries and Excel export
- An installer and a portable build for desktop tools, with automated tests
- Docker deployment, with the code, the accounts and the API keys in your name
Tools we use
- Node.js
- TypeScript
- OpenAI API
- Anthropic API
- Telegram Bot API
- Salla API
- Next.js
- Docker
- Python
- Playwright
- SQLite
- Excel
Typical timeline
3 to 6 weeks for an assistant or automation tool added to an existing business
6 to 10 weeks as part of a full first release.
Salamah Checker was a two-month engagement.
Projects where we did this
- Projects
- 2
Automation, 2025
Salamah Checker
A Windows desktop tool that checks permits in bulk on the Saudi Civil Defense Salamah portal, by permit number, by range or from an Excel file, running up to 30 browser sessions in parallel.
- Python
- Playwright
- SQLite
- Excel
AI assistant
IAI
An Arabic AI customer assistant for a Saudi 3D-printing store on Salla. It answers shoppers on the store's web widget and on Telegram from a searchable knowledge base, with live store data and an admin dashboard.
- Node.js
- OpenAI API
- Anthropic API
- Telegram Bot API
Why clients trust us
You own the code and accounts
The source code repository, the Google Play and App Store developer accounts, and the server, cloud and domain accounts are in your name from day one. We work in them with access you grant, and they stay with you if we stop working together.
An NDA before you explain
We sign a non-disclosure agreement before you explain the idea, not after the first meeting.
30 days of free fixes
For 30 days after launch we fix, free of charge, any bug in what we delivered against the agreed scope. New features, scope changes, and problems caused by third-party services or by someone else's changes are not covered.
Optional monthly support plan
A set number of engineering days each month for fixes, store updates, performance work, keeping an eye on production and new features when needed, with priorities agreed at the start of each month. It is paid, and the price is on request.
Questions about this service
- Questions
- 6
Will the assistant answer things that are not in my store's information?
It is not meant to. It answers from the knowledge base and the store data, and in IAI a confidence threshold decides when to hand the chat to a person instead. You set the hand-over rules in the dashboard.
Do you work with Salla stores?
We have built one assistant connected to a Salla store, IAI, which reads products and orders through the Salla API. We have not built Salla themes or apps for the Salla app store.
Which channels can the assistant run on?
IAI runs on a chat widget in the store and on a Telegram bot. We have not shipped a WhatsApp assistant yet; one can be connected through the WhatsApp Cloud API if you need it.
Which AI model do you use?
IAI uses the OpenAI and Anthropic APIs. The API account and key are in your name, so you pay the model provider for usage directly.
How long does an automation tool take?
We usually plan 3 to 6 weeks as an add-on to your existing work. Salamah Checker, with its 30 parallel sessions and 179 tests, took two months.
Do you have results figures for IAI?
We publish none, because we have no documented usage figures. We would rather say so than show a number we cannot back up.
Related reading
Bulk browser automation with Playwright: 30 parallel sessions in Salamah Checker
How I built a Windows tool that checks permits in bulk on the Salamah portal across up to 30 parallel browser sessions, resumes where it stopped, and retries with backoff.
6 min read
Tell us what you want to build
Message us on WhatsApp or send a short brief, and we reply with a proposed scope and a rough timeline. If you want a rough duration first, try the estimator.