Al-Moayyad Abuljdail

whoami

Al-Moayyad Abuljdail

I work on making models read less and still answer correctly.

Download the CV

About

cat about.md

Al-Moayyad Abuljdail at a KAUST Academy ceremony.

I am a computer science student and research intern working on machine learning, mostly on how models read. My research asks when a model has read enough of a long document to answer, and my competition work has repeatedly turned on finding the one property of a dataset that decides the leaderboard.

King Abdulaziz University, BS Computer Science, 2023 to present. Jeddah.

KAUST Academy, Research Intern, November 2025 to August 2026. Selected among the top 100 of 24,000+ applicants.

Experience

history --roles

11/2025 - 08/2026

KAUST Academy

Research Intern

Selected among the top 100 of 24,000+ applicants for an intensive programme in applied machine learning and deep learning.

2023 - present

King Abdulaziz University

BS, Computer Science

Jeddah, Saudi Arabia.

Skills

ls skills/

languages
  • Python
  • Java
  • Dart
  • SQL
machine learning
  • PyTorch
  • transformers
  • scikit-learn
  • deep learning
  • computer vision
tools
  • Git
  • Flutter
  • pandas
  • NumPy

Research

ls -1 research/

Answer-Convergence Stopping (ACS)

1.6%

Every published stopping signal for long-context reading is learned. ACS reads the frozen model's own answer distribution after each chunk and halts once the belief is confident and has stopped moving. No training, no labels, no tuning.

  • research
  • long context
  • S-NIAH
show how it was wonhide
evidence is herestops here

current answer 1731509 unchanged for 3

gate1731509

You are given a 500-page book and one question: what is the door code? The answer is on page 30. A person reads to page 30, finds it, and closes the book. Today's models read all 500 pages, every time.

Every published stopping signal for long-context reading is learned. ACS reads the frozen model's own answer distribution after each chunk and halts once the belief is confident and has stopped moving. No training, no labels, no tuning.

20 chunks. The needle is 1731509, in chunk 11 or thereabouts.

current answer 1731509, unchanged for 3

Stops before the evidence1.6%

Token reduction, LongBench-v250%

Same, asked gate8.4%

Accuracy retained0.980

Savings captured against oracle60.7%

Preprint on arXiv, 28 September 2026. Poster presented at KAUST Academy poster day, 3 September 2026.

write-uppaper on arXiv

Competitions

ls -1 competitions/

Joint first of 4313 points

Each competition turned on one dataset property rather than on modelling.

Tournament standing

show how it was wonhideChallenge 10.9075

100-species butterfly and moth classification, macro-F1

second place0.9075

public0.9183

train

test

0.54 macro-F1

The test set is grayscale while the training set is colour. Matching them moves the baseline from 0.54 to 0.91.

macro-F1 0.54. Private 0.9075.

macro-F1 0.54

detailswrite-up

show how it was wonhideChallenge 50.6961

Arabic medical question to specialty classification

first place0.6961

public0.6941

  • alefما هو أحدث دواء السكر حالياً في السوق وشكراً.
  • maqsuraماهى الفواكه التى يجب على مرضى السكرى تجنبها
  • ta-marbutaما هى أعراض الغدة الدرقية
  • diacriticsلماذا يكون السكري أقل شيوعاً بين الأطفال؟
matching rows
41
private score
0.5618

41 repeated questions are 40% of both train and test and carry mixed labels. That 40% is a lookup, worth +0.103.

41 matching rows of 3957. Private 0.5618.

41 matching rows, private 0.5618

detailswrite-up

show how it was wonhideChallenge 60.7007

Game genre multi-label from title and description, macro-F1

second place0.7007

public0.7044

0.4587 macro-F1

  • 0.808
  • 0.708
  • 0.093
  • 0.340
  • 0.761
  • 0.112
  • 0.097
  • 0.699
  • 0.199
  • 0.551
  • 0.676

Four of 11 labels sit below 6% prevalence, and a single 0.5 threshold crushes them: on the linear baseline Puzzle scores 0.097 and Arcade 0.093. Giving each label its own threshold lifts macro-F1 from 0.459 to 0.646, more than any modelling change did.

macro-F1 0.4587.

macro-F1 0.4587

detailswrite-up

Projects

ls -1 projects/

  • Timely

    Cross-platform mobile app for event countdowns and scheduled reminders. Dart.

    • Dart
    • Flutter
    • mobile

  • Tadarabon

    Interactive Arabic learning platform with gamified exercises and teacher progress tracking. Built at Eduthon.

    • UI/UX
    • Eduthon
    • education

  • Quua

    Arabic health and fitness tracking app covering weight, steps, calories, and sleep. Dart.

    • Dart
    • Flutter
    • health

  • DSC-ChatBot

    WhatsApp bot automating club operations for Data Science Club KAU, with bilingual FAQ answers.

    • WhatsApp
    • automation
    • bilingual

  • Robocon / SAQER

    Drone assembly and testing for the SAQER competition, including motor and sensor integration.

    • robotics
    • drones
    • SAQER

CV

cat contact.md

Open to research and engineering roles, and to graduate study. The quickest way to reach me is email.

Download the CV[email protected]GitHub