18 years old Founder & CEO, OXIEDO Inventor of ORMAS
Every AI model on earth is a black box. OXIEDO is ending that. Its engine, ORMAS, finds its own broken parts, repairs them mid-training and writes a tamper-evident record of every change, so banks, hospitals and defence teams can finally train AI on the data they have been locked out of. Raadh built it at 18, on one graphics card he paid for himself.
Recognition · 2026
Accepted at DeepMath after double-blind review, a reviewer for NeurIPS 2026, ranked first by the Cosmos Institute and taken into 1752vc Ignite from the top 1% of applicants: ten independent panels, one year.
Cosmos InstituteORMAS was the highest-ranked application in its cycle
1752vc IgniteOXIEDO accepted to the Ignite cohort from the top 1% of applicants
The BridgeFrom a first-round interview with Entrepreneur First in London to the final round of its San Francisco programme
500 GlobalInvited to a first-round interview with the global venture firm on 22 October
FreshmangoOffered a place in the equity-free accelerator after a single interview
IARCO 2026International Academic Research Competition finalist; nearly 500 participants from 60 countries
Cohere LabsOpen Science Community member; Cohere cited his solo work on ORMAS for “remarkable initiative in ML safety and auditability”Rokib Al Dhin Raadh is the 18-year-old founder and CEO of OXIEDO and the inventor of ORMAS, a self-repairing neural network that identifies, repairs and records its own failures while it trains. His stability result for self-repairing training was accepted at DeepMath 2026 after double-blind review. He is a reviewer for the NeurIPS 2026 Trustworthy AI for Good workshop and is a member of Cohere Labs’ Open Science Community.
He grew up in Dhaka and taught himself machine learning. He built ORMAS alone: 16,316 lines of PyTorch and 383 controlled experiments on one RTX 3090, released as an open preprint with code that reproduces every result. OXIEDO, which he founded in 2023, licenses the architecture to regulated industries that need an auditable record of what their models learned. Before ORMAS he built OXIMO, a 40,933-line multi-agent operating system with 2,069 passing tests, and founded five ventures between the ages of twelve and seventeen, selling a stock-prediction system for about $10,000 at fifteen.
OXIEDO licenses ORMAS, the self-repairing neural network Raadh invented, on-premise to regulated industries. Every training run produces a tamper-evident record of what the model learned: evidence for model-risk teams and the EU AI Act’s high-risk obligations.
A 40,933-line multi-agent operating system with 2,069 passing tests. It turns a one-sentence brief into a working organisation and creates the specialist roles it lacks.
How OXIMO works →Removing OXIMO cut output by 91%; restoring it took output above its original level. A twelve-month controlled study on a live UK e-commerce company.
Read the field study →Accepted · Double-blind review · Poster
Raadh’s stability result for ORMAS, the self-repairing neural network, was accepted for poster presentation at DeepMath 2026, the Conference on the Mathematical Theory of Deep Neural Networks. DeepMath takes theory only and reviews it double-blind: specialists judged the proof without knowing who wrote it, or that he is eighteen and self-taught.
This year it is hosted by Ohio State and organised by researchers from Johns Hopkins and Michigan, with invited speakers from Stanford and UPenn; past editions were supported by the National Science Foundation and the Simons Foundation. He presents his poster on 30 October.
“Self-Repair as a Bounded Disturbance: Input-to-State Stability of Neural Network Training Dynamics”
Raadh reviews submissions on mechanistic interpretability, attribution, auditing and post-deployment monitoring, the problems ORMAS was built to address, for the NeurIPS 2026 workshop in Paris.
1752vc placed OXIEDO in its Ignite cohort from the top 1% of applicants, and the equity-free Freshmango programme offered a place after a single interview. Raadh reached the final round of The Bridge (Entrepreneur First, San Francisco) after two interview rounds, and 500 Global has invited him to a first-round interview on 22 October.
Cohere’s research lab welcomed him into its Open Science Community, citing his solo development of ORMAS across 383 experiments.
Selected for the final stage of the International Academic Research Competition, which drew nearly 500 participants from 60 countries.
The full paper and a 36-page supplementary are public on Zenodo (DOI 10.5281/zenodo.21730363), with code that reproduces all 383 experiments, including the 80.3% vs 10.0% recovery after a layer is destroyed.
The ORMAS application ranked highest of every application in its cycle; Cosmos invited it to its technical track.
ORMAS is a self-repairing neural network architecture: it finds its own failing components during training, repairs them without stopping, and logs every repair with its cause. Standard networks cannot say which of their parts failed. A single error signal updates every parameter at once, so interpretability has worked from the outside, reconstructing a finished model’s behaviour after the fact.
ORMAS works from the inside. It bounds each node’s path to the loss at four operations, so every component can be measured while the network trains. Three signals run together: global backpropagation, a per-node local loss through a shared bottleneck, and health-gated self-correction that diagnoses a failing node, repairs it, and logs the repair with its cause.
Raadh’s analysis of that process, adapted from Sontag’s Input-to-State Stability framework, is the first formal local stability result for a self-correcting neural architecture, and was accepted at DeepMath 2026 after double-blind review. The evidence base is 383 controlled experiments across four architectures, every run reproducible from seed.
R. A. D. Raadh
Accepted after double-blind review, DeepMath 2026: Conference on the Mathematical Theory of Deep Neural Networks, Ohio State University, 29–30 Oct 2026. Poster, 30 Oct.
R. A. D. Raadh
Preprint, Zenodo, 1 Aug 2026. Full paper, 36-page supplementary and code reproducing all 383 experiments. DOI 10.5281/zenodo.21730363.
R. A. D. Raadh
Technical report, 2026. Twelve-month controlled field study on a live UK company (Black Bloxie LTD): removing OXIMO cut output by 91%.
UpcomingPoster · DeepMath 2026Ohio State University, Columbus
The stability result behind ORMAS, accepted after double-blind review at the Conference on the Mathematical Theory of Deep Neural Networks.
Conference siteUpcomingTalk · Cohere LabsCohere Labs Open Science Community
Every convergence proof for gradient descent assumes each update is a gradient. Self-repairing training breaks that assumption and still converges. The talk shows why, and exactly where the proof stops.
Event pageReplies are typically sent within one working day.
Questions and opportunities are always welcome.