Home / Services / Practice 02
Generative AI and large language models
Retrieval-augmented assistants trained on your own documents, answering in Modern Standard Arabic and Kuwaiti dialect.
Case: assessment generation
A test-bank assistant built on a curated academic corpus
Prof. Salah Alnajem developed a generative AI assistant that creates test banks and examinations for the courses he teaches at Kuwait University. It is an enterprise-grade intelligent document assistant that implements Retrieval-Augmented Generation, combining OpenAI’s GPT and the Claude API with Azure AI Search to produce answers grounded in a curated knowledge base of academic documents rather than in the model’s general training.
Teaching materials and textbooks are ingested either as digital files or via OCR conversion, then combined with prompt engineering tuned to the assessment task. The result removes most of the manual effort in producing multiple versions of the same question across classes and exam sittings, and frees teaching time for curriculum development and instruction.
What it changed
- Consistency of question quality, style and formatting across every generated assessment.
- On-the-fly oral question generation for students with visual or hearing needs, in one-to-one sessions.
- Real-time streaming responses, semantic caching, and hybrid search combining keyword and vector similarity for retrieval accuracy.
- Deployed as a production web application on Microsoft Azure.
Case: Kuwait International Law School
Two assistants inside a law school
We designed a generative AI chatbot system for Kuwait International Law School and integrated it into the school website and its learning management system, enabling faculty and students to interact in classical Arabic and Kuwaiti dialect.
Admissions assistant
Answers applicants’ frequently asked questions about admission in natural language, trained on the school’s own admissions knowledge. Deployed on the admissions portal page of the college website.
Course-content assistant
Answers students’ questions on the substance of law texts, applied to the content of Explanation of the Kuwaiti Penal Code, Special Section by Dr. Faisal Al-Kandari, professor at Kuwait International Law School.
What we offer
We design intelligent chatbot systems using generative AI and large language models, so that users can interact linguistically with a computer through natural language processing and advanced machine-learning retrieval. In practice this means a system that answers questions about a defined body of knowledge — a set of FAQs, a book, a document archive — in Arabic or Kuwaiti dialect, and returns the answer in natural language rather than a list of links.
Where the material cannot leave your network, the same architecture runs against models hosted inside your own environment.
AI agents in production
Four systems running inside a law school
Each connects to the institution’s own live database, applies the institution’s own rules, and runs on its own servers. The AI interprets and explains the data; it does not invent it.
Academic Intelligence Assistantمساعد الذكاء الأكاديمي
Staff query live student records in plain Arabic or English — no SQL, no report requests, no spreadsheet work. It answers with pass and fail rates by course and term, demographic breakdowns, and full grade histories. A risk-scoring engine rates every student 0–100 on failed courses, attendance failures, incomplete grades and GPA, sorting them into four tiers with specific advisor triggers. It also runs genuine statistics — Pearson correlation, R-squared, p-values, linear regression, ANOVA — and renders scatter plots and trend charts in the chat itself. Student data is read per query and never stored; personal identifiers such as phone numbers and Civil IDs are excluded.
Al-Murshid, the electronic academic advisorالمرشد الإلكتروني
A staff member enters one student number and receives a nine-section advising report in seconds: profile, performance level, GPA standing against the 2.00 threshold in Article 63 of the bylaws, warning and dismissal risk, enrolment duration, recent grades, outstanding failures, and four to six recommendations grounded in that student’s actual record. Every figure is computed from the database by the School’s own regulations, so two students in comparable situations receive comparable assessments regardless of who reviews them. Reports export to PDF in Arabic or English. Retrieval was optimised from 8.3 seconds to under one, so a report can be produced within the span of a conversation.
Course and faculty evaluation analysisمنصة تحليل التقييم
Reads each term’s student evaluation responses directly and turns them into a report management can act on. Positivity is scored per question rather than averaged, making questions and instructors directly comparable. It reads the complete set of free-text comments rather than a sample, surfacing recurring themes only where at least 10% of comments raise them — while flagging serious matters such as harassment, discrimination, safety or academic integrity regardless of how few times they appear. Student names and numbers are stripped on the server before any data leaves the database.
Automated advising and schedulingالإرشاد والجدولة الآلية
Derives each student’s next-term courses from the bylaws — prerequisites, credit caps, failed-course priority, elective pools, graduation catch-up — then builds a clash-free timetable across the cohort. The most recent verified run advised 1,835 students across 420 sections with zero clashes. Advising and scheduling are deterministic: the same inputs always produce the same outputs, and every decision traces to a stated rule or an article of the bylaws. An AI optimisation pass may propose section moves but cannot override any rule — a deterministic validator is the gatekeeper.
Start here
Tell us what your text is doing wrong.
Send us a sample — a set of citizen comments, a document archive, a support inbox — and we will come back with a written read on what is achievable, what it would take, and whether you need us at all.
Send a sample of your text with the form. We reply in writing.