Roshan Mohammed
AI / ML Engineer
Building intelligent systems from GPT architectures to computer vision pipelines.
I like turning research ideas into working software this site walks through what I'm building and why.
> ls ./projects
What I've built
DocuMind
Upload a PDF and chat with it directly. Ask questions and get answers grounded in the document's actual content.
GPT from Scratch
A GPT-style language model built end to end from scratch: tokenizer, attention, and training loop. It learns to predict the next token, so feeding it 'I love' produces a continuation like 'I love Machine Learning.'
PlantNet39
Fine-tuned EfficientNetB2 on the PlantVillage dataset to classify 39 categories: 38 plant disease classes plus a dedicated class for images that aren't leaves at all.
CrossSight
A multimodal image captioning model pairing an EfficientNetB0 encoder with a Transformer decoder. Upload an image and it generates a natural language caption.
Replicating ViT
A from-scratch replication of the Vision Transformer (ViT) paper: patch embeddings, positional encoding, and the transformer encoder, classifying images with no convolutions at all.
> status --current
What I'm building now
I've spent the last stretch shipping projects across NLP, computer vision, and applied ML. Right now I'm pointed at two harder problems.
AI Tutor with a custom 2D rendering engine
I learned complex architectures by implementing them from scratch and pushing AI tools to explain the hard parts but chat alone can't build visual intuition. So I'm building a tutor that shows concepts through a custom 2D rendering engine.
Agentic Plant Health Assistant
Combining PlantNet39 (disease classification), DocuMind (document intelligence), and an agentic reasoning layer to turn a photo of a sick plant into an actionable treatment recommendation for farmers and gardeners.
> cat ./blog
Blog
Writing about what I'm learning while building these projects. Coming soon.
> cat ./about.md
About
I hold a BSc in Computer Science and an MSc in Artificial Intelligence, where I built a foundation in machine learning, deep learning, reinforcement learning, NLP, and data visualization. AI engineering, working hands-on with LLMs, RAG, and agents, is new enough that it isn't really taught yet, so I've been learning it myself by implementing things from scratch: transformers, vision models, and full RAG pipelines, rather than stopping at using pretrained models.
On the data side, I work regularly with Python for analysis: pandas, NumPy, and matplotlib for cleaning, exploring, and visualizing data. I'm currently looking for entry-level opportunities, either as an AI/ML engineer or a data analyst, where I can keep building and learning on real problems.
Education
BSc in Computer Science
MSc in Artificial Intelligence