Hey you, I’m Abdullah.

Problem

20,000 lectures behind look-alike tiles.

Every daily feature was hidden in a three-dot menu, and the player, the screen people used most, was the lowest rated in the audit at 2.1/5.

The challenge was to fix this without breaking the habits 200,000 people had already built.

Research

What people said, and what they actually did.

A bilingual survey ran through the WhatsApp groups and channels where users already were, plus moderated sessions on their own devices, with every task measured against targets agreed in advance.

The survey is self-report and six sessions are directional, so every session finding is either backed up at scale or flagged as unproven.

• ~1,900 survey responses
• 6 moderated sessions
• 3 weeks collection window

Findings

Users didn’t need new features. They needed the existing ones to work better.

Three intents kept surfacing across the survey and the sessions.

• Find. They know what they want. 55% never used Search, because it was hidden.
• Return. They’re mid-series. 60% had saved something they later couldn’t find.
• Explore. No particular goal. The smallest group, and stakeholders chose to prioritise current users.

Solution · Navigation

First, the app needed navigation.

A four-tab bar replaced the kebab, each tab mapped to one behaviour: Home is your activity, Browse is discovery, Library is structured content, Search is direct retrieval.

Home became a dashboard: Recently Played, Favorite Audios, Favorite Series and Downloaded.

Solution · Hardest call

Deciding not to redesign the Library.

Around 200,000 people had already memorised it, so it kept its shape. We only removed the categories the survey showed nobody used, going from 19 to 12.

One taxonomy rule then applied across 20,000+ files: one row per series, variants one level deeper, bilingual titles.

Solution · Player

Playback, rebuilt as two surfaces.

A persistent mini player for the background, and a full-screen player with a real scrub bar and exposed speed control for everything else.

Visible state went on every row: saved, downloaded, or already played.

Results

Shipped, then re-tested on the same tasks.

Eight weeks after release, using the same protocol, against targets set before testing. These come from moderated sessions rather than usage tracking, and Browse and onboarding remain unproven.

• 0% → 92% save-for-later completion
• 0% → 88% “is this saved?” first click
• 4.0 → 0.3 seek errors
• 2.1 → 4.5 player ease, out of 5

Learnings

Good judgment sometimes means doing less design.

The next cycle has two goals: analytics in the product, and closing the accessibility gaps that didn’t make the first release.

• The hardest call was holding back. Letting Browse and Library coexist was harder than designing either one.
• Showing state mattered more than adding actions. The tasks that improved most were the ones where the fix was making things visible.

Problem

An answer you can’t check is a research task.

A lawyer is accountable for every piece of advice they give. Saudi law also changes, so “the law says” is only useful if you know which version said it.

The reframe: not “make the answer more accurate”, but “make every claim cheap to check”.

Approach

A citation is only as good as the three layers under it.

A source link only works if the law is stored in pieces small enough to point at, and if retrieval reliably finds the right piece.

• Structure the law. A legislation CMS for the content team: chapters, parts, articles, annexes and amendments.
• Measure retrieval. An evaluation tool for the AI team: did the right articles come back, and in what order?
• Show the source. Legal chat for lawyers: each claim links to its article, which opens beside the answer.

Solution

Four ways to show a source.

I judged every option by one question: how much work is left for the lawyer between reading a claim and trusting it?

• A link on every claim won. It opens the exact article beside the answer, with the cited text highlighted.
• Then, the version. A correct article in an outdated version is still wrong advice.

Solution · Content

Making the law addressable.

You can only link to an article if the system knows where each article starts and ends. So the content team rebuilds each law as a tree before it’s published.

• Source priority. Tell “the law says” apart from “a guideline suggests”.
• A translation check next to Publish. An unchecked translation is the wrong text.

Solution · Quality

Measuring whether the right article comes back.

A wrong article, shown with confidence, is worse than no citation. Teams upload test sets and compare runs by precision, recall, F1 and nDCG.

The citation a lawyer sees is the output of retrieval, so this was the most direct way to improve what they’d trust.

Results

The claim-to-source pattern shipped.

The source opens as a third panel next to the chat and the contract, so a lawyer can check a claim while drafting without switching tools. Malakah serves 1,000+ legal professionals.

Learnings

Trust is a path, not a badge.

Next, I’d test citation density on long answers, measured by time to verify.

• The visible pattern depends on invisible structure. The CMS and the evaluation tool are what made the citation link true.
• Design for the people behind the AI too. The content and AI teams were users of this system as much as the lawyers.

Problem

Every share was a copy nobody could take back.

Drafts went back and forth by email, finance approved in WhatsApp, and 9 of 12 people had sent an outdated version at least once.

Moving contracts into one workspace only helps if sharing inside it is as easy as forwarding an email, and much safer.

Solution · Model

Four levels, each answering one question.

Rather than one long list of permissions, each level answers a single question, in the place where people ask it.

• Organisation. Who is on our team? Admin or Member.
• Project. Who works on this matter? Edit, Query or View.
• Contract. Who can open this document? Per person, per file.
• Signing. Who signs, in what order, until when? With an audit trail.

Solution · Key decision

A role for asking, not editing.

Reading a 60-page supply agreement isn’t what a finance approver wants. With only Edit and View, they’d get edit rights they shouldn’t have, or no way to ask.

Query lets them run questions across a project’s files without changing anything.

Solution · Sharing

Sharing that stays visible and reversible.

The access list sits inside the share window, so you never grant access without seeing who already has it.

• One click to change. Revoking access happens where it was granted.
• Pending is its own state. Invited people stay greyed out, and can be withdrawn before they get in.

Solution · Interface

The interface reflects your access.

Unavailable actions are removed, not disabled. A viewer never meets a button that fails, so the menu itself tells them what their access is.

Ownership is labelled with an Owner badge, plus separate “My projects” and “Shared with me” views.

Solution · Signing

Every signature leaves a record.

Signing involves people outside the company. Access for outside parties can expire, and the audit trail records every view and signature and exports as a certificate.

Results & learnings

AI creates new kinds of access.

One permission model shipped across organisation, project, contract and signing, used by Malakah’s 1,000+ legal professionals.

• Describe the action, not the role. A one-line “what you can do” beat any role name, especially across two languages.
• Show access where it’s granted. It did more for safe sharing than a separate settings page would have.

Problem

A capable studio with a brand that undersold it.

The first brand book stopped describing the business as the company grew.

• A story out of step. It positioned a multi-industry studio as an AI legal platform.
• A category-default palette. Violet and navy on white, shared with countless SaaS brands.
• A voice without rules. Adjectives gave no guidance on writing a headline or an error message.

Solution

Engineered Clarity.

Every decision follows one test: does it make Stixor clearer? The hexagonal mark kept its geometry, so recognition isn’t lost.

• Three colours with a job each. Charcoal 50, Ivory 40, Orange 10. Orange is never a background.
• Two type families. Helvetica Neue for structure, DM Mono for precision.

Solution · Voice

We don’t just claim things work. We explain how.

Three pillars, with rewrites for the moments where tone matters most.

• Precise, not pedantic. “Automate your manual data entry with 99% accuracy.”
• Confident & direct. “Connection failed. Please check your network or refresh the page.”
• Human, but professional. “Welcome. Let’s get your workspace set up in three steps.”

Results

Less noise. More signal. Same Stixor.

A brand that matches the work: calm, exact and recognisably its own, with rules simple enough for the whole team to use.