← All reports

AI-Powered Exam Generation & Automated Grading

How fine-tuning a specialized 1.7B model and integrating OpenCV automated test generation and grading.

How fine-tuning a specialized 1.7B model and integrating OpenCV automated test generation and grading.

Support agent

Knowledge agent

Client

AI-Powered Arabic/English MCQ Platform

Industry

Team

Lead AI & Full-Stack Engineer

Build

2025 – 2026

Sheet

C-02

FIG. C-02 — Interactive Exam Editor & OMR Optical Grading System

FIG. C-02 — Interactive Exam Editor & OMR Optical Grading System

Scale 1 : 1

01

Survey — The problem

Manual exam drafting and paper grading burdened educators with immense overhead.

Manual exam drafting and paper grading burdened educators with immense overhead.

Built an end-to-end educational assessment platform designed to eliminate the friction in exam creation and grading.

Technical Achievements

  • Bilingual LLM Fine-Tuning: Developed a specialized 1.7B Arabic/English LLM for MCQ generation, fine-tuning Qwen on 80,000 in-house Arabic/English question-answer examples using SFT, DPO, ORPO, and RLAIF.

  • Automated OMR Grading: Integrated an Optical Mark Recognition (OMR) pipeline using OpenCV to automatically scan and grade bubble sheets exported from the platform.

  • Full-Stack Architecture: Interactive exam editor, PDF text extraction, and administrative dashboard built with Next.js and Django REST Framework (DRF).

  • Accolades: Awarded 1st place among 25+ projects at the University Graduation Project Exhibition.

Calls / week

1,900

Missed at peak

1 in 5

People on it

6 receptionists

02

Blueprint — What we built

One front desk, three exits.

One front desk, three exits.

01 · InputPatient call + chat1,900 / week
02 · AgentFront-desk agentListen · Verify · Book · Confirm
03a · OutputAppointment bookedAuto · 74%
03b · OutputReminder sentAuto · 74%
03c · ReviewNurse callbackClinical · 26%
1.2 s Avg. decision
01 · InputPatient call + chat1,900 / week
02 · AgentFront-desk agentListen · Verify · Book · Confirm
03a · OutputAppointment bookedAuto · 74%
03b · OutputReminder sentAuto · 74%
03c · ReviewNurse callbackClinical · 26%
1.2 s Avg. decision

03

Result — Measured after 90 days

0
0

missed calls since go-live

+0%
+0%

bookings per week

0.0 / 5
0.0 / 5

patient satisfaction

“Patients get an answer on the first ring, and my team finally has time for the people standing at the desk.”

“Patients get an answer on the first ring, and my team finally has time for the people standing at the desk.”

“Patients get an answer on the first ring, and my team finally has time for the people standing at the desk.”

Ines Vermeer

Practice Manager, Halcyon Clinics

Next report

→

Kiln Studio — 71% of tickets auto-resolved.

Kiln Studio — 71% of tickets auto-resolved.

Kiln Studio — 71% of tickets auto-resolved.

Let’s build
something useful.

Have a software engineering opportunity or an AI application to discuss? Reach me at akilalsharafi@gmail.com. Based in Ibb, Yemen; working in Arabic and English.

Output · 01

Email Akil

↗

Let’s build
something useful.

Have a software engineering opportunity or an AI application to discuss? Reach me at akilalsharafi@gmail.com. Based in Ibb, Yemen; working in Arabic and English.

Output · 01

Email Akil

↗

Let’s build
something useful.

Have a software engineering opportunity or an AI application to discuss? Reach me at akilalsharafi@gmail.com. Based in Ibb, Yemen; working in Arabic and English.

Output · 01

Email Akil

↗

Let’s build
something useful.

Have a software engineering opportunity or an AI application to discuss? Reach me at akilalsharafi@gmail.com. Based in Ibb, Yemen; working in Arabic and English.

Output · 01

Email Akil

↗