I'm Mahroz — a computer science student building LLM pipelines, data platforms, and full-stack apps. Currently a web developer at GAOTEK.
From backend APIs to LLM retrieval pipelines to pixel-perfect interfaces.
The problem behind my flagship project: finding an answer buried in a document shouldn't take longer than asking for it.
Six projects, six different problems — language, data, design, and real-time systems.
One taught me retrieval, the other taught me sequence modeling — different problems, same instinct to go past the tutorial.
Reading through long PDFs to find one answer is slow. I built a system that parses any document, breaks it into meaningful chunks, and lets you ask it questions in plain English — with the model grounding every answer in the actual text instead of guessing. The hard part wasn't the model call; it was everything around it: parsing messy documents cleanly, chunking without losing context, and keeping retrieved passages small enough to fit the context window without losing the answer.
A deep learning project exploring sequence modeling for music generation. I trained LSTM networks on MIDI piano performances so the model could learn the temporal patterns underneath a piece — which notes tend to follow which, how phrases resolve, how timing holds together — and then generate original musical sequences from what it learned. The interesting part was framing music as a language problem: notes and durations as tokens, a performance as a sentence, and generation as sampling the next most plausible token, one step at a time.
Open to internships, collaborations, and interesting problems in AI/ML and full-stack development.