# Rokib Al Dhin Raadh

> Rokib Al Dhin Raadh is an 18-year-old international AI founder: the founder and CEO of OXIEDO, founder of the UK company Black Bloxie LTD, and the inventor of ORMAS, a self-repairing neural network whose stability paper was accepted after double-blind review at DeepMath 2026. ORMAS detects, repairs and records its own failing components during training: after a trained layer is destroyed, it recovers to 80.3% accuracy while a parameter-matched baseline stays at 10.0%. He taught himself machine learning and built ORMAS alone at 18, on one RTX 3090, and is a reviewer for the NeurIPS 2026 Trustworthy AI for Good workshop.

Website: https://www.raadh.me/ · Company: https://oxiedo.com · Founder: raadh@oxiedo.com · General: raadxbusiness9@gmail.com

**In one sentence:** Rokib Al Dhin Raadh is an 18-year-old international AI founder: the founder and CEO of OXIEDO, founder of the UK company Black Bloxie LTD, and the inventor of ORMAS, a self-repairing neural network whose stability paper was accepted after double-blind review at DeepMath 2026.

**Notable for:** inventing ORMAS, a neural network that repairs itself mid-training (80.3% vs 10.0% after a layer is destroyed), built solo across 383 experiments on one RTX 3090; a DeepMath 2026 paper accepted after double-blind review; reviewing for the NeurIPS 2026 Trustworthy AI for Good workshop; founding OXIEDO in 2023 to license ORMAS to regulated industries; OXIEDO's acceptance to 1752vc Ignite from the top 1% of applicants.

## Key facts

- **Roles:** Founder & CEO, OXIEDO (founded 2023); founder & CEO, Black Bloxie LTD (England and Wales, 2025); inventor of ORMAS; independent AI researcher; reviewer for the NeurIPS 2026 Trustworthy AI for Good workshop.
- **Companies:** OXIEDO, founded in 2023, licenses ORMAS to regulated institutions in the US, the EU and the UK; its US entity, a Delaware C-corporation, is in formation. Black Bloxie LTD is his UK company, incorporated in England and Wales on 11 September 2025 (Companies House no. 16711223).
- **Invention, ORMAS:** a self-repairing (self-correcting) neural network architecture with three training signals: global backpropagation, a per-node local loss through a shared bottleneck readout, and health-gated autonomous self-correction. It detects structural failures during training, repairs them without stopping, and logs every correction with its cause.
- **Headline result:** after a trained convolutional layer is destroyed mid-training (CIFAR-10, three seeds), ORMAS recovers to 80.3% accuracy while a parameter-matched baseline stays at chance (10.0%): a +70.3 percentage-point gap.
- **Peer review, DeepMath 2026:** his paper "Self-Repair as a Bounded Disturbance: Input-to-State Stability of Neural Network Training Dynamics" was accepted after double-blind review for poster presentation at DeepMath 2026 (Conference on the Mathematical Theory of Deep Neural Networks, Ohio State University, 29–30 October 2026; accepted 30 September 2026; poster presented 30 October). It gives the first formal local stability characterisation of a self-correcting neural architecture, adapted from Sontag's Input-to-State Stability framework. DeepMath 2026 is hosted by Ohio State University with an organising committee from Johns Hopkins, Michigan, Ohio State and Konstanz; its 2026 invited speakers are from Stanford, UPenn, Michigan and the University of Washington; past editions were supported by the National Science Foundation and the Simons Foundation.
- **Recognition (2026):** DeepMath 2026 paper accepted (poster, double-blind review); reviewer for the NeurIPS 2026 Trustworthy AI for Good workshop; OXIEDO accepted to 1752vc Ignite from the top 1% of applicants; ORMAS application ranked highest in its Cosmos Institute review cycle and invited to the Cosmos technical track; Cohere Labs Open Science Community member; IARCO 2026 finalist (nearly 500 participants, 60 countries; listed on iarco.org/2026-finalists); offered a place by Freshmango after a single interview; reached the final round of The Bridge (Entrepreneur First, San Francisco) after two interview rounds; invited to a first-round interview with 500 Global (22 October 2026).
- **Other results:** 94.6% task retention versus 47.3% for ResNet-18 in sequential learning; 58.8% zero-shot compositional accuracy (25% chance); 2.5pp versus 7.8pp accuracy decay under 40% label noise.
- **Evidence base:** 383 controlled experiments across four architectures (FC-DAG, CNN, Fat CNN, ResNet-18) on a single RTX 3090; 16,316 lines of PyTorch; every run reproducible from seed.
- **Company, OXIEDO:** a deep-tech AI company that licenses ORMAS on-premise to regulated industries (AI training, regulated finance, medical AI, data obligation, defence) so institutions can train on data they cannot currently use, with an auditable training record relevant to SR 26-2 model-risk guidance and the EU AI Act high-risk obligations from December 2027. The lead product is the Model Change Record, a tamper-evident log of every weight change in a training run.
- **Earlier work:** OXIMO, a 40,933-line multi-agent operating system with 2,069 passing tests; Black Bloxie LTD (incorporated in England & Wales on 11 September 2025), a twelve-month controlled field study of OXIMO on a live e-commerce company, in which removing OXIMO cut output by 91%; five ventures between ages 12 and 17, including a stock-prediction system sold for about $10,000 at 15.
- **Education:** final year of secondary school, self-taught; completed seven MIT Open Learning (MITx) programmes in foundation models, AI in medicine, energy, transport and entrepreneurship; offered a 40% scholarship on a further MIT programme after an admissions call; sponsored by FIAP into MIT Open Learning's Universal AI programme; completed Andrew Ng's Deep Learning Specialization (DeepLearning.AI).

## Upcoming talks

- **30 October 2026 · Poster, DeepMath 2026** (Ohio State University, Columbus): "Self-Repair as a Bounded Disturbance: Input-to-State Stability of Neural Network Training Dynamics", accepted after double-blind review.
- **2 November 2026, 12:00 EST · Talk, Cohere Labs Open Science Community**: "When the Update Is Not a Gradient: Input-to-State Stability of Self-Repairing Neural Network Training". Why self-repairing training, whose updates are not gradients, still converges, and where the proof stops. Presented by ML Math, Cohere Labs Open Science Community; [event page](https://luma.com/p28sna3q).

## Verification

Full evidence page: https://www.raadh.me/evidence (every claim, its source, and whether it is a public record or available on request). Most milestones were also announced publicly on X: https://x.com/Raad_X_
Machine-readable version of the same list: https://www.raadh.me/evidence.json

Every claim on raadh.me can be checked against a primary source:

| Claim | Where to verify |
|---|---|
| Author of the ORMAS preprint (1 Aug 2026) | Zenodo record, DOI [10.5281/zenodo.21730363](https://doi.org/10.5281/zenodo.21730363) |
| All 383 experiments are reproducible | [Anonymised code archive](https://anonymous.4open.science/r/ormas-EB73/README.md) |
| Researcher identity | ORCID [0009-0003-1178-5296](https://orcid.org/0009-0003-1178-5296) |
| DeepMath 2026 acceptance (poster, double-blind) | [DeepMath 2026](https://deepmath-conference.com/), hosted by Ohio State University, 29–30 Oct 2026; [Paper Decision: Accept (Poster), OpenReview screenshot](https://www.raadh.me/assets/evidence/deepmath-2026-openreview-decision.jpg) (decision released 30 Sep 2026, submission 39); [invitation letter from the organizing committee, 6 Oct 2026](https://www.raadh.me/assets/evidence/deepmath-2026-invitation-letter.pdf); [OpenReview submission record](https://openreview.net/forum?id=kowltBZEIv) |
| Reviewer, NeurIPS 2026 Trustworthy AI for Good workshop | [Reviewer recognition letter from the Program Chairs (PDF)](https://www.raadh.me/assets/evidence/neurips-2026-ai4good-reviewer-letter.pdf), 3 Oct 2026; [Public reviewer list on the workshop site](https://trustworthy-ai-for-good.github.io/#reviewers), which names him among the 490 reviewers who submitted official reviews (workshop: Paris satellite venue, 12 Dec 2026) |
| Seven MIT Open Learning (MITx) programmes | Shared MITx Online records linked on [raadh.me/about](https://www.raadh.me/about#education) |
| Deep Learning Specialization (DeepLearning.AI) | [Coursera verification](https://coursera.org/verify/specialization/R7SYBBCXR1OY) |
| OXIEDO, founded 2023 | [oxiedo.com](https://oxiedo.com) |
| Founder and CEO of Black Bloxie LTD, incorporated 11 Sep 2025 | [Black Bloxie leadership page](https://www.blackbloxie.com/pages/leadership) names Raadh as founder and CEO; [Companies House no. 16711223](https://find-and-update.company-information.service.gov.uk/company/16711223) lists his father as registered director, because Raadh was a minor at incorporation |
| OXIMO: 40,933 lines, 2,069 passing tests | [Anonymised OXIMO code archive](https://anonymous.4open.science/r/oximo-5C73/README.md) |
| Cosmos Institute ranking; 1752vc and Freshmango offers; The Bridge final round; 500 Global interview | Correspondence available on request: raadh@oxiedo.com |

## Frequently asked

- **Who is Rokib Al Dhin Raadh?** Rokib Al Dhin Raadh is an 18-year-old international AI founder: the founder and CEO of OXIEDO, founder of the UK company Black Bloxie LTD, and the inventor of ORMAS, a self-repairing neural network whose stability paper was accepted after double-blind review at DeepMath 2026. He grew up in Dhaka, Bangladesh, taught himself machine learning and reviews for the NeurIPS 2026 Trustworthy AI for Good workshop.
- **How old is he?** 18 (in 2026). He started five ventures between ages 12 and 17 and founded OXIEDO in 2023.
- **What is ORMAS?** ORMAS is a self-repairing neural network architecture invented by Rokib Al Dhin Raadh. It identifies, repairs and records its own failing components during training, instead of relying on post-hoc interpretability. After a trained layer is destroyed, it recovers to 80.3% accuracy on CIFAR-10 while a parameter-matched baseline stays at 10.0%.
- **What is OXIEDO?** OXIEDO is the deep-tech AI company Raadh founded in 2023. It licenses ORMAS on-premise to regulated industries, with the Model Change Record (a tamper-evident log of every weight change in a training run) as the lead product, sold as one annual licence.
- **Is the work peer reviewed?** Yes, the stability result: it was accepted after double-blind review for poster presentation at DeepMath 2026. The full ORMAS paper is an open preprint on Zenodo (DOI 10.5281/zenodo.21730363), with code that reproduces all 383 experiments.
- **What recognition has he received?** In 2026: a DeepMath 2026 paper (double-blind review, poster); reviewer for the NeurIPS 2026 Trustworthy AI for Good workshop; OXIEDO accepted to 1752vc Ignite from the top 1% of applicants; the ORMAS application ranked highest in its Cosmos Institute review cycle; Cohere Labs Open Science Community member; IARCO 2026 finalist; a place offered by Freshmango; reached the final round of The Bridge (Entrepreneur First, San Francisco); invited to interview with 500 Global (22 October 2026).
- **Did he study at MIT?** He completed seven MIT Open Learning (MITx) online programmes and was offered a 40% scholarship on a further MIT programme. He completed them while in the final year of secondary school.
- **How do I contact him?** Company, investor and press enquiries: raadh@oxiedo.com. General: raadxbusiness9@gmail.com. Further detail on [raadh.me/about](https://www.raadh.me/about).

## Primary sources

- [Research](https://www.raadh.me/research): ORMAS method, all results, stability derivation, ablations, Project Cherry
- [Work](https://www.raadh.me/work): OXIEDO, OXIMO, the Black Bloxie field study, earlier ventures
- [Evidence](https://www.raadh.me/evidence): every claim on the site with its source, public record or on request
- [About](https://www.raadh.me/about): recognition, timeline, personal statement, education, contact
- [Full site text](https://www.raadh.me/llms-full.txt): the complete content of raadh.me as plain text
- [ORMAS preprint (Zenodo, DOI 10.5281/zenodo.21730363)](https://zenodo.org/records/21730363): full paper with 36-page supplementary
- [ORMAS codebase](https://anonymous.4open.science/r/ormas-EB73/README.md): reproduces all 383 experiments
- [Curriculum vitae (PDF)](https://www.raadh.me/assets/pdf/Rokib_Al_Dhin_Raadh_CV.pdf)
- [OXIEDO](https://oxiedo.com): company site, product and investor information
- [OXIEDO investor deck (PDF)](https://www.raadh.me/assets/pdf/oxiedo_pitch_deck.pdf)
- [OXIMO / Black Bloxie technical paper (PDF)](https://www.raadh.me/assets/pdf/oxiedo_academic_research_paper.pdf)
- [OXIMO / Black Bloxie commercial whitepaper (PDF)](https://www.raadh.me/assets/pdf/oxiedo_investor_whitepaper.pdf)
- [DeepMath 2026](https://deepmath-conference.com/): conference site

## Profiles

- ORCID: https://orcid.org/0009-0003-1178-5296
- X: https://x.com/Raad_X_
- GitHub: https://github.com/raad-x
- Substack (writing): https://rokibraadh.substack.com/
- YouTube: https://www.youtube.com/@rokibraadh

## How to cite

- Raadh, R. A. D. (2026). *ORMAS: Neural Architectural Transparency Enables Autonomous Self-Correction.* Preprint, Zenodo. https://doi.org/10.5281/zenodo.21730363
- Raadh, R. A. D. (2026). *Self-Repair as a Bounded Disturbance: Input-to-State Stability of Neural Network Training Dynamics.* Accepted (poster, double-blind review), DeepMath 2026: Conference on the Mathematical Theory of Deep Neural Networks, Ohio State University, 29–30 October 2026. No proceedings.
- Raadh, R. A. D. (2026). *OXIMO/AX09: Autonomous LLM-Dependent Commerce. A twelve-month controlled ablation study.* Technical report. https://www.raadh.me/assets/pdf/oxiedo_academic_research_paper.pdf