Mohammadhossein R.

I build machine learning and generative AI systems end to end, from messy raw data to a deployed, monitored service and I optimize for the metric that actually matters to your business, not just accuracy or AUC. Recent work includes: • A self-hosted, multimodal LLM platform (70B-class model served via vLLM, speech-to-text/text-to-speech, image understanding, and a 4-tier safety/guardrail pipeline) running entirely on private infrastructure — no third-party API costs, no data leaving the client's servers. • A multilingual RAG system ingesting PDFs, scans, images, and audio into Qdrant, with hybrid retrieval, reranking, and full source/page citation. • A real-time traffic-quality model for a large consumer platform, rebuilt around actual profit (not a proxy metric), cutting revenue loss while rejecting a quarter of low-value traffic. • A 7-day student attendance early-warning model built from 12.8M raw records, benchmarked across 16 classifiers and tuned to match a school's real intervention capacity. Stack: Python, PyTorch, XGBoost/LightGBM, Optuna, Hugging Face, vLLM, LiteLLM, Qdrant, FastAPI, Docker, AWS, GCP. I work directly with founders and technical leads, clear scoping, regular check-ins, and code/documentation handed off cleanly at the end.