Hi, I'm Mohammed Saqlain

Qwen-Docs Finetuned is a Qwen3-VL-8B vision-language model optimized for document understanding and structured markdown generation from images such as scanned pages, PDFs, screenshots, and technical documents. Fine-tuned using Unsloth and Hugging Face TRL, it preserves document layout, supports headings, lists, tables, and inline formatting, and is suitable for technical and academic content. The model enables low-VRAM inference (4-bit quantized) and is ideal for extracting structured information from complex documents.

Syndata is a synthetic data generation platform for creating Q&A pairs from PDF documents. It is useful for evaluation tasks and data generation, enabling users to quickly generate high-quality question-answer datasets from their own documents.

Built a fully automated content pipeline that generates and uploads shortform videos using LLMs and YouTube API. Integrated GitHub Actions to schedule daily creation and publishing workflows for consistent automation and automated the generation of video scripts, titles, hashtags, and metadata for optimized reach.

Developed a tool to transform web content into presentation videos with synchronized narration and music. Automated slide creation, narration, and audio generation using Gemini and text-to-speech models and streamlined content curation through web scraping and AI-driven summarization for presentation flow.

Built a command-line LLM assistant to manage and modify calendar events via natural language. Integrated an Agentic RAG framework for reasoning over contextual information about scheduled events and automated the discovery and scheduling of upcoming Codeforces contests based on user intent.

Fine-tuned multilingual BERT for Turkish hate speech classification using parameter-efficient LoRA adapters. Implemented token-level explainability using SHAP to visualize word contributions in predictions and built a rewriting pipeline to paraphrase hateful text while retaining tone and semantic meaning.