Mohamed Hassani

I create and solve software engineering tasks used to train and evaluate language models. The work is close to code review, but it also requires designing the situation that the model has to understand.

• Designed tasks covering debugging, code generation, refactoring, optimization, testing, API development, and algorithms.
• Wrote and reviewed Python, JavaScript, Java, C++, and Swift code across different architectures and programming styles.
• Compared generated implementations for correctness, security, performance, maintainability, readability, and missing requirements.
• Rewrote incomplete solutions and explained the edge cases, performance issues, and reasoning errors that made them fail.
• Created realistic developer scenarios and reviewed other contributors' work before submission.

ArtFusion built interactive kiosks rented by exhibition clients. Visitors took photos, created personalized portraits, received a branded print on site, and accessed a digital copy through a QR link after completing the client form. I owned the technical system behind that experience.

• Defined the hardware setup with an embedded Linux workstation, NVIDIA GPU, touchscreen, cameras, and thermal printer.
• Built the main PyQt6 application for touch and hand gesture interaction, with configurable content and visual identity for each exhibition.
• Designed the image generation pipeline with Stable Diffusion, ControlNet, and LoRA, reaching generation times under 30 seconds in difficult conditions such as low light and group photos.
• Developed backend services for sessions, telemetry, forms, collections, user data, content delivery, and QR access.
• Optimized memory and inference, integrated the Linux hardware, and handled deployment, testing, and on site reliability.

repository

Worked on data quality projects for language model training, where the useful output was not just an annotation but a reliable example that another model could learn from.

• Validated large datasets against detailed annotation guidelines as the project requirements evolved.
• Reviewed generated text for accuracy, consistency, usefulness, and errors that could affect later training.
• Investigated ambiguous cases, explained the problem, and escalated issues when the written guidance was not enough.
• Worked with a distributed team while keeping the review process consistent across many examples.

Owned multiple machine learning projects from data collection and preparation through model development, training, deployment, and API integration. The work covered object detection, OCR, document understanding, visual search, image generation, and the annotation tools and data pipelines that connected them.

• detection and segmentationBuilt and deployed computer vision systems for object detection and segmentation. Trained DETR and YOLO model variants, prepared the training data, and converted models for faster inference in applications.
• document aiBuilt invoice and receipt extraction from input formatting through text detection and structured document understanding. Compared the results with AWS Textract and prepared the inference path for batch processing and service use.
• search and data collectionBuilt an image search and retrieval system over marketplace data, including collection, metadata lookup, image representation, indexing, and similarity search. Built scraping pipelines with multiple workers, retries, rate limits, and deduplication across different sources.
• services and annotation toolsTurned models into inference services and connected them to APIs. Maintained annotation workflows and built frontend and backend tooling for segmentation data, including mask encoding that could move between Python, JavaScript, and TypeScript.
• generative visionWorked on image restoration, image captioning, content moderation, Stable Diffusion experiments, and a trained image generation model packaged for serving.

• Designed data pipelines for cleaning, validation, augmentation, annotation, and dataset conversion before model training.
• Deployed trained models as reusable inference services and connected them to the applications around them.
• Worked across Python, Node.js, and TypeScript when the machine learning system crossed into product tooling.

Built a conversational AI assistant for mental health support as my master's end of studies project. The work moved from finding usable dialogue data to training models, designing conversation flows, and deploying a web application and API.

• data collectionBuilt scraping pipelines with Selenium and Beautiful Soup to collect conversation examples from public websites, forums, Reddit, support communities, and other online sources. I also explored extracting useful material from mental health podcasts.
• model experimentsCleaned, normalized, and deduplicated the data before experimenting with several generative language models. Fine tuned GPT 2 and DialoGPT with Hugging Face Transformers for supportive dialogue.
• conversation and deploymentCombined generative responses with structured conversation flows in Rasa, then deployed the chat application and REST API with Flask on Google Cloud Platform.

• Connected web data collection, dataset preparation, model training, conversation control, and deployment in one project.
• Built a working application rather than stopping at a language model experiment.