Maram Ali

Profile summary

Education

Studying computing and AI

Career highlights

Professional experience

I finished a university internship around Machine Learning and Natural Language Processing, plus some speech-to-text stuff, kind of. I worked on checking and validating how accurate the JSON text comes out from audio calls, using Omatel data. Also, I built a Whisper-based ground truth pipeline to turn the call audio into really accurate transcripts, and in the same stretch, I prepared datasets for fine-tuning a Mistral LLM.

• •Developed a Whisper-based ground-truth pipeline for accuracy+ call recordings, achieving 33.3% WER. Prepared instruction-based samples and fine-tuned Mistral for transcript correction, improving accuracy by 20% over the baseline.
• •Prepared a supervised fine-tuning dataset for the Mistral LLM containing precision by 98%+ transcript samples, using incorrect transcripts as inputs and ground-truth transcripts as outputs for automated spelling and grammar correction.
• •Analyzed Mistral model performance using WER, CER, BLEU/ROUGE, or exact-match accuracy, depending on the evaluation methodology.

Key skills

Projects

This project focuses on developing a question-answering (QA) system that interacts with users through the Furhat robot, aiming for a natural and human-like conversational experience. The system allows users to ask questions verbally or through text, and Furhat responds with both spoken answers and corresponding facial expressions and gestures.

Objective: Developed a real-time, mobile-based translation system designed to bridge communication gaps between deaf or hard-of-hearing individuals and public service providers using Arabic Sign Language (ArSL). Target Platform & Architecture: Built as a high-performance native Android application utilizing an MVVM (Model-View-ViewModel) software architecture with a clean separation across UI, ViewModel, and Repository/Service layers. As the UI/UX Designer, I spearheaded the design and engineered user-centric interfaces, guaranteeing exceptional usability and robust cross-platform compatibility. Key Performance Metrics & Accessibility -Accuracy & Latency: Engineered for low-latency mobile inference with a target model accuracy threshold of 70% or higher, ensuring rapid response times to maintain natural conversation flow during public service interactions. -User Adoption & Accessibility: Designed with compliance to WCAG 2.1 (Level AA) accessibility standards, featuring customizable light and dark themes, adjustable text sizing, high-contrast visual layouts, and intuitive onboarding flows tailored for both deaf users and government service employees.

NahwiFix is a modern Arabic grammar and punctuation checker for Android, built with Kotlin and Jetpack Compose; there is also a website version. Features: -Instant Arabic Grammar & Syntax Checking: Detects and highlights errors in real time, including Hamzat Al-Wasl and Al-Qat', Tanween Al-Fath, subject-verb gender agreement, and accusative/genitive cases. -Arabic Punctuation Rules: Corrects Arabic commas (،), Arabic question marks (؟), spacing around conjunctions (و), and misplaced punctuation. -Detailed Correction Inspector: Inspects every issue with the original error snippet, suggested replacement, grammatical reason, and authoritative rule reference. -One-Click Corrections: Apply corrections individually or apply all suggested fixes simultaneously. -Local Persistence with Room: Stores past check history securely and offline using the Room Database. - Full Localization & RTL Support: Native Right-To-Left (RTL) Arabic layout with an English bilingual toggle. Final Benchmark Results Summary: -Linguistic Domain / Subsection (Precision (P): 99.2%; Recall (R): 98.5%; F (0.5) Score: 99.0%) Technical & Orthographic (Spelling/Omissions)

Certifications

Languages