Mohammad Takneshan

During my undergraduate studies, I worked as a part-time Artificial Intelligence researcher at the University of Debrecen under the supervision of Dr. Attila László Gilányi, specializing in computer vision research with a focus on AI-driven markerless motion capture technology. The research was accredited and funded by the Talent UD program, presented at the University of Debrecen student research conference, and received a nomination to Hungary's national research conference; the thesis paper is currently being prepared for publication in an academic journal. The implications of my research can improve the field of motion capture by reducing setup time by up to 80% and overall costs by 50% through optimized, camera-based AI workflows, lowering production time by 40% and post-processing requirements by 30% using improved data processing and motion reconstruction techniques, and advancing the accessibility and fidelity of motion capture for applications in animation, biomechanics, sports analytics, and virtual production.

I did a remote web development internship at a service-based company in Oslo, Norway during my university break. I worked with other interns and experienced colleagues on planning, designing, and building web services for clients, using React.js, Tailwind CSS, and shadcn/ui. It was a great experience learning to collaborate in a real dev environment and build clean, responsive UIs.

I was referred to a Machine Learning Engineer role at 4ty-engine, a startup business where I collaborated with another engineer to develop an LLM-powered chatbot for a student engagement application. The chatbot was designed to help students find relevant resources, answer questions, and connect with peers more effectively. I focused on prompt engineering, retrieval-augmented generation, response quality, and inference optimization to help students find resources, answer questions, and connect with peers.

I helped prospective IB students in completing their Internal Assessment, a research project part of their Diploma Program. Reviewing and offering academic guidance on formatting and structuring their academic papers. My areas of expertise included helping the student in, English, Business Management, Mathematics, and Physics investigations and papers. I enjoyed guiding IB students, helping them improve their research and submit stronger papers.

I built Travel Avenue to give Osaka Group a direct sales channel it owned rather than rented. Selling through Amazon and Noon means the marketplace sets the terms and keeps the customer relationship; this is the company's own storefront for suitcases, bags, and travel accessories in the UAE, and the first time it sold directly. It's a custom Shopify theme with a Cloudflare Workers backend. I rebuilt the collection taxonomy, filtering, and localised URLs so customers can find what they want, and designed the campaign, sale, and cart experience around how people shop for luggage. The site runs in English and Arabic, with the translation files checked in CI so a missing string fails the build rather than reaching a customer. The Worker also serves the bilingual shopping assistant, moderated article comments, and audit events, while PostHog covers event tracking, behavioural analytics, session insights, and web vitals, so changes are judged on real customer behaviour rather than opinion.

A responsive site reaches someone once; an installed app reaches repeat customers on the device they actually shop from, and lets the brand stand on its own rather than as one listing among competitors. So I extended the direct channel onto mobile with a fully native app for iOS and Android. It shares the same Shopify backend and Cloudflare Worker as the website but has its own native UI, built in Expo and React Native with TypeScript over a typed data layer on Shopify's GraphQL Storefront and Customer Account APIs, so it pulls catalogue, cart, and customer data straight from Shopify and hands checkout back to Shopify's hosted flow. Sign-in uses Customer Account OAuth with PKCE and the session lives in the device Keychain. On top of that sit the things only a native app can do: App Intents and Siri Shortcuts, home-screen widgets, deep links, haptics, an optional Face ID lock, and an accessibility settings screen for text size, motion, and contrast, in English and Arabic.

Customer data was heading into a per-seat SaaS CRM, which meant paying per head, accepting vendor lock-in, and keeping our customer records on someone else's platform. I deployed and customised the open-source Twenty CRM on the company's own AWS infrastructure instead, so it owns its customer data outright and pays no per-seat fee. I extended it with a custom object model so sales, support, and operations work from one shared view of the customer: B2B leads with score, priority, and status, warranty cases, Shopify orders, website events, opportunities, notes, tasks, and dashboards. Shopify stays the commerce source of truth, so the sync service writes through the Twenty API rather than the database, with dry-run, idempotency, and cursor safety. It runs on Docker Compose on AWS, with internal access behind a deny-by-default WorkOS gateway providing role-based access, MFA, and audit logging, and secrets brokered through AWS Secrets Manager.

Masaryk University is the second largest university in the Czech Republic and as an assigned representative, I built and maintain the website for the Masaryk International Medical Students Association so its committee can run events, publish study materials, and issue membership cards without needing a developer. I am also responsible for buying and managing the domain and providing ongoing technical support. It's a student-focused platform: a public side with events and an upcoming-events timeline, an authenticated member portal, and a Study Hub where students find study materials organised by subject and year. Behind it there's an admin content-management portal so the committee can manage events, resources, and members themselves, plus a membership-card generator that produces PDF cards. The stack is Next.js and React, Convex for the backend and database, Clerk for authentication, Tailwind CSS for styling, and Framer Motion for interactions, deployed on Vercel with cleanly separated development and production environments.

Checking product listings, search performance, and indexing meant a person clicking through Merchant Center, Search Console, and Drive dashboards by hand. I built a read-only Model Context Protocol server in Python with FastMCP so AI agents can investigate all three directly, through OAuth 2.0 credentials, with no path to changing production data. Read-only is the point: the tools expose catalogue health, search performance, indexing, sitemaps, and file analysis, and nothing can write, so an agent can dig into an SEO or listing problem without any risk to the live catalogue.

As part of my journey for learning German, I'm developing an German learning application combining interactive lessons with courses, featuring adaptive curricula, vocabulary quizzes, and speaking/writing/grammar exercises. includes dynamic difficulty adjustment based on user performance, exam preparation aligned with Goethe Institute standards, and collaborated with a German tutor for content accuracy.

Developed a multimodal emotion recognition (MER) system as my BSc thesis at the University of Debrecen, supervised by Lecturer Róbert Lakatos. The system fuses four modalities (text via RoBERTa, audio via Wav2Vec2, video via CNN+Transformer, and facial expressions via FacialNet) through attention-based fusion that dynamically weights each modality based on the emotional content of the input. Trained and evaluated on the MELD dataset, the four-modality system achieved 71% emotion recognition accuracy, approaching the 75% human baseline, and 85% on 3-class sentiment analysis , outperforming all unimodal configurations. A cross-lingual evaluation on a Persian dataset revealed a 47% drop in text model accuracy, empirically demonstrating that cultural and linguistic context is a primary barrier to deploying emotion AI globally , a finding with direct implications for real-world multimodal AI systems. Also built a full-stack desktop application with real-time training monitoring and LLM-powered result analysis, and conducted an industry assessment covering healthcare, e-learning, and conversational AI applications.

Developed a deep ensemble system for Type 2 diabetes prediction combining neural networks, CNNs, and AutoML, achieving 80% accuracy and 0.89 ROC AUC.

I developed a full-stack web application for sentiment analysis using natural language processing (NLP), in collaboration with a classmate, who worked on the frontend. I focused on backend development, implementing model training, deployment, and real-time inference. Improved app efficiency by 40% by developing a multilingual sentiment analysis system using Python, Hugging Face Transformers, and NLP tokenization for emotional tone classification; enhanced accessibility and user satisfaction through a responsive, multilingual UI (HTML, CSS, JavaScript) that received 80% positive user feedback; and enabled real-time sentiment insights with 3× faster processing by optimizing deep learning models using GPU-accelerated PyTorch.

I developed an autonomous car-parking system in Unity as part of a reinforcement learning course, using ML-Agents and PyTorch for training and simulation. The project's reinforcement learning approach demonstrated clear improvements over traditional rule-based methods: reduced training time by 40% through efficient environment design and RL workflow optimization, achieved 20% faster convergence by tuning PPO hyperparameters via YAML configuration improving training stability, and reached 98% validation accuracy across over 50% of training epochs using TensorBoard for monitoring and ONNX for model deployment.