About
ABI AI exists to help organisations put AI into real use — by building the systems themselves, and by producing the human data that makes those systems work.
Purpose
A great deal of AI work stalls between demonstration and deployment. A model performs well in a notebook, and then meets real data, real users, and real accountability. ABI AI works in that gap.
We do two things. We build AI products and custom systems for organisations that need something fitted to their own workflow. And we help AI companies design and operate the human-data programmes — collection, annotation, evaluation, red-teaming — that determine how good their models actually become.
ABI AI was established in 2026. We are a small, senior operation, and we scope engagements to what we can genuinely deliver rather than to what sounds impressive.
Company identity
ABI AI is a licensed free-zone company in Dubai. The details below are provided in full for due-diligence purposes.
Founder
ABI AI was founded by Soroush Farghadani, who works at the intersection of applied machine-learning research and large-scale human-data operations — two capabilities usually found in separate people.
He leads human-data work at xAI, a frontier AI laboratory, where he sets direction for frontier data and evaluation programmes. Before that he led the development of AI for autonomous drone navigation in GPS-denied environments at SOTI’s advanced aerospace research group.
That combination is the reason ABI AI does both things it does: the same person who knows what a frontier lab needs from its human data has also shipped machine-learning systems that had to hold up outside the conditions they were trained on.
A note on how to read this section. The roles and research below are Soroush Farghadani’s own professional and academic experience. They are not presented as projects delivered by ABI AI. Where ABI AI is the contracting party, that is stated explicitly.
Apr 2025 – Present · Palo Alto, CA (remote)
Sets direction for frontier data and evaluation workstreams at a frontier AI laboratory — deciding focus areas, defining quality standards, and turning broad model-assessment needs into executable plans for distributed expert teams.
Current · Contracted and delivered through ABI AI
Leads human-data operations for G2i, a current ABI AI customer. This is the one engagement on this page where ABI AI is the contracting party.
May 2022 – Mar 2025 · Toronto, Canada
Led end-to-end development of AI systems for autonomous drone navigation in GPS-denied environments — the kind of problem where a model that only works in benign conditions is worse than none at all.
Sep 2021 – May 2022 · Advisor: Prof. John C. Hull
Designed and tuned a novel distributional reinforcement-learning method for optimal hedging under uncertainty in financial markets, published and presented at academic and industry venues.
Oct 2019 – Aug 2021 · Munich, Germany (remote)
Published applied machine-learning research spanning interpretable graph neural networks and medical imaging, including deep-learning pipelines for disease prediction and early dementia detection under limited-data constraints.
xAI and SOTI are referenced solely as part of the founder’s employment history. Neither is a customer, partner, or affiliate of ABI AI.
Founder publications
Research co-authored by Soroush Farghadani, including joint first authorship on two papers. Authors are listed in published order and every entry links to the publisher or arXiv record, so each citation can be checked directly against the source.
J. Cao, J. Chen, S. Farghadani, J. Hull, Z. Poulos, Z. Wang, J. Yuan
Frontiers in Artificial Intelligence, 2023
Reinforcement learning with quantile regression applied to hedging a stochastically arriving derivatives book. Also presented at RiskMinds and CDI.
A. Kazi, S. Farghadani, I. Aganj, N. Navab
Machine Learning in Medical Imaging (MLMI), MICCAI, 2023
Joint first authors — the paper marks these authors as contributing equally.
An attention module that exposes which input features drive a clinical prediction, rather than returning an unexplained score.
R. Shirkavand, S. Ayromlou, S. Farghadani, M. Tahaei, F. Pourakpour, et al.
arXiv preprint, 2021
Joint first authors — the paper marks these authors as contributing equally.
Learning from limited, weakly labelled clinical imaging data.
Customers
ABI AI currently supports G2i’s human-data operations, with founder Soroush Farghadani serving as Head of Human Data.
This is the company’s current named customer engagement. We do not publish client counts, case-study metrics, or testimonials we cannot substantiate.
We are happy to provide further documentation for due-diligence or onboarding purposes.
