eTROC: an AI agent marketplace companion to buy and sell

eTROC is an AI agent marketplace for Morocco: you tell an agent what you want to buy or sell in your own words, on WhatsApp, Telegram, iMessage, email or the web, and it searches live listings near you, compares them, watches for new matches and drafts every message for your approval.

This showcase covers what ImadDhin built for eTROC, why an agent sits at the centre of the product, how listings and location work, and the engineering decisions behind a marketplace that has to load fast on a mid-range phone and stay trustworthy inside a chat thread. It does not publish traffic or conversion figures; the product has just launched in its first market and those numbers will come later.

Why buying and selling second-hand needs a companion

In Morocco, buying and selling is spread across Facebook Marketplace groups, Avito, Moteur.ma for vehicles, Jumia for new goods with delivery, and countless WhatsApp conversations. A buyer looking for a used phone or a family car opens four apps, repeats the same search, scrolls past duplicates and stale posts, and then negotiates in a chat thread with no context. A seller posts the same item in three places and answers the same questions all day.

The problem is not a missing listing page. It is the work between wanting something and closing a deal: searching, comparing, asking the right questions, following up, and noticing when something better appears. That work is repetitive, conversational and personal, which is exactly where an agent earns its place.

eTROC therefore keeps a complete marketplace, with listings, photos, saved items, an inbox and a seller dashboard, and puts an agent on top of it that works wherever the user already talks.

What does the eTROC agent do for buyers?

The agent is a marketplace concierge working for one person. It receives a request in plain language, in Darija written in Arabic or Latin letters, French, Arabic or English, and replies in the same language and script. A typical request sounds like a message to a friend: a used iPhone in Casablanca under a budget, a washing machine that can be delivered, a first car for a student.

• Search live listings by need, budget, condition and area, then present numbered results with price, condition, city and a link
• Widen the search once when results are thin, by dropping the city or relaxing the price, before saying nothing matched
• Open the full details of any result and compare options side by side
• Import a listing link pasted from another marketplace so it can be compared with eTROC results
• Save listings to the user's Saved list on eTROC
• Draft a message to the seller, show the exact text and recipient, and send it only after an explicit yes
• Create alerts that watch for new listings matching a search and notify the user on their connected apps

One rule shapes every reply: the agent may only present listings returned by its tools in that conversation. It never invents a price, a seller or availability. When eTROC has few matches, it suggests which marketplace usually carries that kind of item and invites the user to paste links from it, so the comparison still happens in one place.

How does eTROC help people who sell?

Selling starts with a two-step listing: photos and details first, then a location. Location is not decoration; it powers matching. When a listing goes live, eTROC checks it against buyers who are looking for something similar in that area and notifies them on the web and on their connected apps as soon as it appears.

Sellers get a dashboard with their listings, a match inbox that shows interested buyers, and one inbox for conversations. When a buyer writes, the agent can summarise the thread and suggest a reply, and the seller decides what goes out. Messages from third parties are treated as content to summarise, never as instructions the agent should follow, which closes an obvious route for manipulation.

Account types keep the interface honest. Buyers use eTROC for free and never see seller tools they do not need. Seller and Business accounts unlock listing management, stock clearance tools and the match inbox.

One agent across WhatsApp, Telegram, iMessage, email and the web

People in Morocco already buy and sell in chat. eTROC treats each messaging app as an entry point into the same agent rather than a separate bot. A user connects WhatsApp, Telegram, iMessage or Gmail from their account settings, and the agent keeps one memory of the conversation, saved items and alerts across all of them.

Each channel gets its own writing style. WhatsApp replies are plain text with short lines and at most five results. iMessage replies are shorter still. Email replies read like a short email with a greeting. In the web app, Discover is a split view with listings on the left and the agent on the right, and the agent can use links and short paragraphs.

Sending anything on the user's behalf always takes two steps: the agent shows the exact text and recipient, and it acts only after an explicit yes in the same conversation. The pattern is covered in more depth in human-in-the-loop AI agents, and it matters even more on a marketplace, where a careless message can commit someone to a price.

Live listings near you from the first day

An agent is only as useful as the inventory it can search, and an empty marketplace is the classic cold-start problem. eTROC launched in Morocco with live listings from partner marketplaces alongside listings posted directly on eTROC.

Vehicles come from Moteur.ma and new goods with delivery come from Jumia Maroc. A scheduled worker reads their public listing pages a few pages at a time, respects each site's robots.txt rules, identifies itself with a named crawler and a contact address, and stores every item with its source and original link. Listings that stop appearing are archived automatically, so stale posts do not linger in results.

Imported listings are labelled with their source. Buyers open them on the original marketplace, and eTROC does not message those sellers on their behalf. They are not submitted to search engines as eTROC pages either; only listings posted on eTROC are indexed.

Location ties it together. On the first visit eTROC asks for the user's location and, if they decline, falls back to Casablanca. Results are filtered by distance with a radius the user controls, and items that ship nationwide appear for anyone shopping in the country.

How listings are matched

Matching combines distance, price fit, intent and meaning. Listings are embedded as vectors when they are created or change, so a search for a family car within a budget can match a listing that never uses those words. Spatial queries handle the radius and the nationwide case in the same database, which keeps search results and agent answers consistent.

The engineering behind eTROC

The product is split into a web app and an agent worker that share one Postgres database, so a listing saved in WhatsApp is already in the Saved list on the web.

Running the agent on Cloudflare Workers keeps each user's agent state durable and close to them, while Hyperdrive avoids opening a new database connection for every message. The web app follows the same patterns we use for Next.js on Vercel builds: server components for listing pages, structured data for search engines, and a mobile-first performance budget measured on mid-range phones.

Analytics respect consent. Measurement starts only after the user accepts the cookie banner, and conversations that begin on WhatsApp or Telegram are recorded server-side as leads without the content of the messages.

Which shortcuts did we refuse?

Some shortcuts were tempting during the build, and each was rejected because it would trade trust for short-term growth.

• No invented or demo listings in production; an empty search is better than a fake result
• No unofficial automation of personal messaging accounts
• No payments inside chat; payments happen in person or through the seller, and the agent never asks for card numbers, passwords or codes
• No messages to sellers on other marketplaces in eTROC's name
• No unexplained ranking; results show the facts that matter, price, condition, city and distance

Buyers are rightly cautious about scams on classifieds. Every one of these rules removes a way for the product to lose that trust.

Rolling out country by country

eTROC is built to expand one market at a time. Each country has its own currency, the local marketplaces the agent can work with, partner sources for live listings, and guidance for the agent about how people buy there. Morocco is the first market, and the structure for the next ones is already in the code, so adding a country is configuration and partnerships rather than a rebuild.

Users can also switch the country they are shopping in by asking the agent, which matters for people who buy across borders or are about to move.

What this means if you are building an agent product

eTROC is a useful reference for any team putting an agent in front of a catalogue, a booking system or a service marketplace. Three lessons carry over directly.

• Ground every answer in tools and real data, and let the agent say plainly when it has nothing
• Meet users in the channels they already use, but keep one memory and one set of rules across them
• Put an explicit confirmation in front of every action that speaks for the user

If you are planning something similar, our AI agent development work covers agent design, tools and evaluation, and AI product development covers the marketplace, data and launch around it. You can book a 30-minute call to walk through your idea, or try the agent yourself on eTROC.