Researchers who ship real software

Iman Saberi is a machine learning engineer and researcher with a Ph.D. in Computer Science from the University of British Columbia and more than seven years of experience taking AI from research into production. His recent work is all about large language models: training them, making them fast and affordable to run, and making sure their answers can be trusted.

As a Data Scientist at Global Relay, Iman trains and fine-tunes LLMs ranging from 1.5B to 30B parameters on multi-GPU clusters of NVIDIA B200s and B300s. His toolkit includes distributed training, preference alignment with DPO, and continual fine-tuning that adds new knowledge without erasing what a model already knows. He also runs the MLOps side: CI/CD pipelines on Kubernetes with monitoring, drift detection and automatic rollbacks, so models stay reliable after launch.

Much of his day-to-day work is inference optimization. Using quantization (GPTQ and NVFP4), smarter batching and KV-cache tuning, he has cut model footprints by up to 74% and made inference more than three times faster. Every change is checked with benchmarks, A/B tests and load tests before it reaches users.

Before that, at Charli Capital, Iman built FISCAL, a system that detects hallucinations and unsupported claims in financial text by checking them against the source documents. Fine-tuning an open-weight LLM for the task raised fact-checking accuracy from 87% to 93%, and the work was published at NeurIPS 2025. He also built document layout understanding pipelines that turn earnings reports, filings and contracts into structured data.

His research on code intelligence, efficient fine-tuning and knowledge-graph retrieval has been published at IEEE SANER, MSR and EMSE, and received an IEEE TCSE Distinguished Paper Award. He also holds a Master’s degree in Computer Science from the University of Tehran.