hi, i'm Muhammad Maroof
Researching core AI architectures and engineering the systems that deploy them.
30+
projects on GitHub
2
research papers
Top 400
on Kaggle
Document Q&A, semantic search, grounded and cited answers
QLoRA, quantization and llama.cpp for hardware without a GPU
FastAPI and Django REST services with PostgreSQL and Docker
Vision, text and audio models in PyTorch
LLM & RAG · 2026
Phi-4-mini-instruct (3.8B) fine-tuned with QLoRA on the QuerySmith-Spider-Bird dataset (derived from Spider and BIRD), then compressed with a partial-residual quantization scheme and served with llama.cpp. The final model reaches 0.743 exact-match accuracy at 22.3 ms/token and 3.5 GB, against an FP32 baseline of 0.713 at 148 ms/token and 14.9 GB (6.6x faster, 4.2x smaller).
Full-Stack · 2026
AI-powered system built around Laya, a structured decision model for fast probabilistic decisions. Instead of calling a generative LLM for every task, Laya turns an unstructured request into structured decisions (category, priority, risk, escalation probability). A deterministic policy engine interprets them and a workflow engine acts (assign, escalate or queue). Example: "Our production database is unavailable and all API requests are failing" becomes Category Technical, Priority Critical, Escalation 0.94, routed to the Engineering team and escalated. Components: Laya decision layer, decision normalizer, policy engine, workflow engine, FastAPI backend, persistence layer, testing layer. Goal: a small specialized model for structured decisions, deterministic software for business rules, and large generative models only when language generation is really needed.
Full-Stack · 2026
Full-stack platform doing syntax validation, bug detection, security scanning, complexity analysis, optimization suggestions and automatic documentation, through a six-stage sequential analysis pipeline with real-time NDJSON streaming.