Atlas App

A ranking that explains itself, criterion by criterion

Most teacher directories sort by signup date or by who paid for a subscription. Atlas App does the opposite: the subject is a FILTER — a teacher who does not teach it never appears — and everything else is a score out of 100 points, spread across eight weighted criteria.

Level carries 30 points, because an excellent middle-school teacher is not a final-year teacher. Area carries 20: distance is the number one reason families abandon home tutoring. Then come budget (14), in-person or online format (12), availability (10), experience (8), profile verification (4) and a free trial lesson (2).

Two rules make all the difference in practice. First, a criterion the family left blank penalises nobody: it drops out of the calculation instead of producing an arbitrary ranking from a deliberately short form. Second, every result carries its reasons and its caveats in plain words — "Teaches the requested level", "Works in Agdal", "Has not declared this level". A recommendation you cannot justify inspires no trust, and a family that does not understand the ranking phones the centre to redo it by hand.

• The subject filters; the seven other criteria score — never the reverse
• A blank criterion is an ignored criterion, not a penalising one
• An incomplete profile scores half rather than being excluded — otherwise new teachers stay invisible
• Online lessons cancel the area constraint, and the score says so
• Results appear BEFORE the family is asked for a phone number
• The browser never receives teacher contact details — ranking is computed server-side

Everything families used to phone about, on one page

A tutoring centre spends its days answering the same four questions: did he show up, what do I still owe, how is he doing, when is the next payment due. The Atlas App parent portal answers all four without anyone picking up the phone.

Parents log in and see their child's real attendance — justified and unjustified absences kept apart, last attendance dated — the exact state of their instalments (total, discount, paid, outstanding, unpaid months), monthly report cards downloadable as PDF, the handouts assigned to them, their library loans with return dates, their timetable and their teachers' comments.

The detail that actually changes behaviour: every deadline carries its own countdown. Not one counter on the instalment, but one per deadline — the instalment, the library subscription, a book return. Those are precisely the dates people forget because they never see them.

• Attendance: justified and unjustified kept apart, last attendance dated
• Finances: total, discount, paid, outstanding, number of unpaid months, currency
• Monthly report cards as PDF, generated on demand and never stored in the clear
• Assigned handouts, downloadable from the same page
• Library: books read, subscription, current loans and return dates
• One countdown per deadline, not a single global counter
• Payment receipts as PDF, without going through the front desk

One review, one verified phone number

A teacher directory is worth exactly what its reviews are worth. Reviews on Atlas App cannot simply be dropped in: the parent enters their name, rating, comment and Moroccan number, receives a six-digit code on WhatsApp, and the review is published only once that code is validated.

The safeguards live in the code, not in a policy page: five attempts maximum, an expiring code, one verified review per number-and-teacher pair, HTML stripped from comments, a ten-character minimum, and a rate limiter on the write route.

Finally, the two rating systems are linked. A verified review also counts as a vote in the teacher's main score — the one shown at the top of their profile and driving the homepage ranking. Without that link, the same profile displayed two different averages in two places, the kind of inconsistency that costs more than having no rating at all.

• Six-digit code sent over WhatsApp before publication
• Five attempts maximum, then a new code is required
• One verified review per number and per teacher
• The average is recomputed from the votes, never incremented — a counter drifts, a recomputation does not
• Verified reviews and star ratings feed the same average

The Moroccan baccalaureate calculator, with the real coefficients

Every spring, the most searched query among Moroccan high-school students is how to compute their baccalaureate average. The sites that capture it are blogs: they answer, but they cannot offer a lesson afterwards. Atlas App answers AND offers.

The official formula is applied as is — 25 % continuous assessment, 25 % regional exam, 50 % national exam — across all seven tracks, each with its own national-exam coefficients. That is exactly the point students miss: the same subject can weigh 9 in one track and 3 in another, and many split their revision time blindly.

The scale is coded in one single place in the product. A calculator and a simulator returning two different results for the same student would destroy trust in the tool in a single visit.

• 7 tracks covered, with per-track national-exam coefficients
• The first-year regional exam is included: a quarter of the final mark, which many learn too late
• A single set of constants shared by every screen
• Three addresses redirect to the canonical page (calculator, simulator, average calculation)

The CV reads itself; the candidate retypes nothing

A centre that recruits loses half its candidates on the application form. Atlas App reverses the order: the teacher uploads a CV, the server reads it, and the form arrives already filled in — identity, highest degree, institution, years of experience, subjects, city, and even the photo from the CV when there is one. The candidate corrects instead of typing.

The extracted text does more than pre-fill: it is indexed. On the centre's side, a deep search finds "agrégation" or "mission française" inside the body of the CV even when no form field mentioned it. A file uploaded by a candidate is deleted from disk as soon as reading finishes, successfully or not.

• Upload the CV, the server reads it, the form is pre-filled — then a human corrects it
• The CV photo is recovered when present
• CV text is indexed: the centre's deep search reads the document, not just the fields
• The file is wiped from disk immediately after reading, including on failure
• A dedicated rate limiter on the parsing route: it is expensive, so it is protected

Local pages built from the teachers who actually exist

A matching platform is not found through a search page: nobody types "teacher search", people type "maths tutor in Agdal". So Atlas App builds its local pages from the profiles actually published: every subject × city pair covered by at least one teacher becomes a "private {subject} lessons in {city}" address, with the matching teachers already ranked on it.

The reverse rule matters just as much: a combination nobody covers does not produce an empty page, it redirects. Serving pages with no real content degrades the whole site's ranking — the classic trap of mass-generated directories.

On top of that: one page per purchase-intent query (comparisons against foreign software, flagship modules, guides), prices published in the open where the entire market hides behind a form, and a seasonal back-to-school page serving all three audiences — families, teachers, centre owners — and linking them together.

• One page per subject × city pair actually covered, never more
• Uncovered combination: a redirect, not an empty page
• Trilingual sitemap generated on the fly, with hreflang alternates and x-default
• Old addresses kept as 301s: dropping an indexed URL means losing what it earned
• Schema.org markup (breadcrumb, FAQ, organisation) emitted with the page, not bolted on afterwards
• Prices published openly — no Moroccan competitor does this

Three languages, three indexes — not a toggle

Plenty of "multilingual" Moroccan sites switch language with a button and keep a single address. To a search engine, such a site has one version: the other two languages do not exist.

Atlas App mounts every one of its pages under a language prefix. Each page therefore exists at three indexable addresses, each declaring its counterparts through hreflang, with an x-default. The language is read from the URL rather than from the browser: that is what lets the server produce the right version immediately, whereas a browser-stored preference does not exist yet when the server replies — and never exists at all for a crawler.

Arabic is not a translation laid over a Latin layout: the document flips to right-to-left, typography included, and every direction-sensitive component — arrows, timelines, sliders — flips with it.

• French, Arabic and English, each at its own indexable address
• hreflang fr-MA / ar-MA / en plus x-default on every sitemap entry
• Language read from the URL, not the browser — the precondition for correct server rendering
• Full right-to-left Arabic, directional components included
• Old language-less addresses redirected with a 301 to their French counterpart

Server-rendered, code-split, cached in front

A matching platform that takes three seconds to show its first line loses the family before its first result. Atlas App is server-rendered: the visitor receives the HTML of the requested page, immediately readable, and only that page's own JavaScript is hydrated afterwards.

That split is not cosmetic. Statically imported, the site's pages shipped as a single 984 kB bundle the visitor downloaded in full just to read the homepage; the production budget refused to compile it at all. Every page is now loaded on demand.

In front of the rendering server, an nginx cache serves already-computed pages with validity windows tuned per address family, and keeps serving the last known version if the application restarts. In production, rendering runs as a two-instance cluster.

• Full server rendering: content readable before JavaScript arrives
• One page, one on-demand chunk (the single 984 kB bundle is gone)
• nginx cache in front of rendering, per-family validity, stale-while-restarting
• Two clustered instances, restart without downtime
• A reveal animation never hides content for long: three safety nets, including a hard time limit