Hi, I'm

Maram
Almaohgra

مرام المهجرع

Product Engineer · Arabic AI & Search

I turn complex technology into simple, useful products.

View my work

Hover the wings · tap to scatter

السهل الممتنع

Simple to use. Hard to make.

About me

Computer engineering and psychology graduate. My philosophy is simple, useful things: products that feel obvious because someone did the hard thinking first.

Product thinking

ضArabic AI

User first

Builder

Psychology lens

Team player

Tools I use

PyPython

ReReact

JSNode.js

SQLSQL

GitGit & GitHub

SMSageMaker

OSOpenSearch

AsAstro

MiMiro

NoNotion

Selected work

Launched

Arabic meaning search

Find a word by describing what it means. Rebuilt as RAG: from 9s+ to under 3s.

Read case study

EmbeddingsOpenSearchSageMaker

Prototype

InPocket

AI receipt scanning for field-service companies. Spec, pricing and launch plan.

Read case study

PRDPricingGo-to-market

Notebook

Food segmentation

Labeling every pixel of a meal photo with DeepLabv3+. mIoU rose from ~63 to ~84.

Deep learningDeepLabv3+Colab

Live

My Arabic blog

Right-to-left, dark mode, and live on every push. Built to stay cheap and mine.

AstroMarkdownGitHub Actions

Experience

Jul – Dec 2025

AI Engineer

Arabic dictionary platform

Owned the semantic search system end to end.

Aug – Oct 2026

Product Management

CorePro · Istanbul

120-hour program, capstone at Demo Day.

How I work

Listen

Learn the problem from the people who have it.

Define success

Agree on what a good result looks like.

Build an MVP

Ship the smallest version that tests the idea.

Measure

Watch what the data and users show.

Decide

Build it for real, or change the solution.

Let's build something simple and useful.

Riyadh · Istanbul · Remote

All work

Case study · Shipped

Searching an Arabic dictionary by meaning

A RAG-based search that helps writers and translators find the word for a meaning they have in mind.

RoleAI Engineer

Team2 engineers, linguists, a designer

WhenJul – Dec 2025

StackPython · SageMaker · OpenSearch

1,000+Searches in the first week (approx.)

9s+ → <3sTypical response time (informal testing)

3,500 → 100Lines sent to the LLM per query (approx.)

User queryA meaning, in Arabic

EmbeddingThe query becomes a vector

Vector searchClosest 100 titles

LLMPicks the best match, returns IDs only

ResultA real chapter and entries

The problem

Writers, translators and researchers sometimes know exactly what they mean, but not the word for it.

A regular dictionary

Needs the word before you can look it up.

A general chatbot

Can answer, but may invent a word that doesn't exist.

This search

Every result comes from the dictionary's own entries.

The linguistics team confirmed the gap, and language specialists said solving it matters.

My role

The first version sent every chapter and entry title to an LLM on each query. It was too slow to be practical, so I proposed rebuilding it as RAG and owned the decisions behind it.

Defined what a good result looks like with a linguist, and used our lists of search terms and expected results to compare options.
Made every result a real entry: the LLM returns only IDs in a fixed format, and the text comes from the database. Inputs are checked for length and malicious code.
Covered the full-stack work: search filters, the onboarding tour and the contact flow.
Proposed smart suggestions that show the closest match instead of "no related result".

Three decisions

An open-source Arabic embedding model, not an API

Tested three models on the linguists' word pairs and picked the most accurate. Hosting it on SageMaker took an outside API out of the request path and gave a fixed running cost.

Gave upEasy setup

Embed titles and chapter names, not explanations

Titles carry a chapter's meaning; long explanations could blur it. Results were already good, so we launched on time.

Gave upMore testing before launch

Query processing in the backend

SageMaker only hosts the embedding model, on the smallest instance that worked, to keep cost low.

Gave upEverything in one managed service

What I'd test next

The LLM only sees the closest 100 results. If the right chapter isn't in that list, it can't pick it. I'd run my test searches through vector search alone and record where the correct chapter lands.

Top 100Top 200Top 300

Most within 100: improve the LLM prompt. Many further down: send more results, or change what gets embedded.

Honest label

Real product with real users. No formal benchmark dataset: quality was judged with the linguists' lists and review. Usage figures are approximate, response times were measured informally, and cost savings are expected, not measured.

Full case study (PDF) Next: InPocket

All work

Product case study · Course project

InPocket

A tool for field-service companies that reads receipts with AI, so they stop losing them and stop typing them in by hand.

TeamSquad of 5, no fixed titles

ProgramCorePro, 120 hours

OutputPrototype

The problem

Receipts get lost, retyped by hand, or guessed. Each of those hurts someone different.

The accountant

Enters every receipt into the system by hand.

The field worker

Pays out of pocket when a receipt is lost.

The company

Keeps records built on guesses and typos.

My role

We decided most things together. I took part in every stage; these are the parts I worked on most.

ProductMVP scope, PRD with acceptance criteria, backlog, Now / Next / Later roadmap, where AI should and shouldn't be used

MetricsExperiment spec, success and failure in numbers

PricingBusiness model, cost per seat, price tiers, pricing test

LaunchGo-to-market plan with an Arabic sales script, readiness checklist, FAQ, growth plan

ResearchRan a user interview

DesignHelped design the MVP wireframes

Three decisions

Scanner first

The full idea also tracked field visits. We built only the scanner to prove one thing first, and left out accounting exports until we knew which systems our customers used.

Gave upVisit tracking and exports, for now

AI reads the receipt; a simple rule picks the category

AI reads merchant, amount and date. A lookup table assigns one of five categories; unknown merchants are picked by hand. Free, easy to explain, and it never makes up an answer.

Gave upSmarter AI categorization

A targeted launch, not a public one

Reach 5 businesses through accountant intros and sales calls, backed by an email, a blog post and a help article.

Gave upA full public campaign

What I'd test next

Before building it for real, I'd interview 10 field-service managers. Is this a real problem for them, and would they pay to solve it?

5 of 10 say yes to bothFewer: don't build it as it is

Honest label

A course project. Nothing worked for real and no real users tried it. I'm not fully happy with it: we never had clear positioning, and we tried to combine two hard problems without one story tying them together. Next time I'd pick one problem and build the whole story around it.

Full case study (PDF) Next: Arabic meaning search

MaramAlmaohgra

About me

Tools I use

Selected work

Arabic meaning search

InPocket

Food segmentation

My Arabic blog

Experience

AI Engineer

Product Management

How I work

Listen

Define success

Build an MVP

Measure

Decide

Let's build something simple and useful.

Searching an Arabic dictionary by meaning

The problem

My role

Three decisions

An open-source Arabic embedding model, not an API

Embed titles and chapter names, not explanations

Query processing in the backend

What I'd test next

Honest label

InPocket

The problem

My role

Three decisions

Scanner first

AI reads the receipt; a simple rule picks the category

A targeted launch, not a public one

What I'd test next

Honest label

Maram
Almaohgra