M.Sc. candidate researching reliability-aware compression of small language models for on-device function calling. I have over 1.5 years of experience shipping LLM features in a production Python backend. My work uses careful evaluation with held-out tests, multiple seeds, and standard significance checks.
Multilingual WhatsApp customer assistant in the Umsiyeti ecosystem. Wired the WhatsApp Cloud API to LLM function calling, with backend workflows that turn AI-generated actions into real reads/writes on a live Supabase / PostgreSQL service database.
Product-discovery agent over a deliberately corrupted catalog. Strict separation between LLM semantic scoring and deterministic validation of every fact — 266 tests, audited by mutation testing. Built as a case study for E-Trink Global.
AI-powered system for analyzing tree leaves and diagnosing plant health, with real-time monitoring and detailed reports for optimal tree care.
Dual-role marketplace for event services in Mauritania — clients discover and book venues, photographers, halls and transport; providers manage listings and orders. Built with React Native (Expo), Supabase and TypeScript, fully multilingual (Arabic / English / French) with RTL. I contributed as part of the team on AI features and backend integration.
A Flutter app that uses AI for enhanced contact management and profile creation, with real-time updates from Firebase. Features smart contact suggestions, detailed profiles, seamless navigation and advanced search.
A platform I built to collect and structure Hassaniya Arabic data — a low-resource Arabic variety with little available training data — to support dialect adaptation and LoRA fine-tuning.
Provider-side application for a delivery service, built with Flutter. Currently in development.
I fine-tuned Qwen2.5-1.5B-Instruct with QLoRA on 4,258 training examples. The fine-tuned model used 18.46M trainable parameters (1.18% of the model) and ran on a single NVIDIA T4 (24.8 min, 3.64 GB peak). On 271 held-out examples, exact-match accuracy rose from 8.58% to 53.73% — a 45.15-point (526% relative) increase. Results were stable across three seeds, and adapters and code are public on Hugging Face and GitHub.
I study how INT8 and INT4 quantization affects function-calling reliability on edge devices with intermittent connections. I also test grammar-constrained decoding and reliability-aware QLoRA as ways to recover accuracy lost to compression.
LoRA fine-tuning to adapt a language model to Hassaniya Arabic, a low-resource Arabic variety with minimal available training data.
Built the Python / FastAPI backend, database, REST APIs, and AI features of Umsiyeti, a digital-services product on the Apple App Store. I worked on backend workflows, database integration, and external API integrations for AI features. I handled model integration, debugging, and production delivery.
Mentored international students in Software Engineering, Cybersecurity, and Artificial Intelligence. I helped them understand technical concepts and organise their learning.
Reliability-Aware Compression of Small Language Models for On-Device Function Calling under Intermittent Connectivity
Kocaeli University · 2025 – 2027. Full detail under Research ↑