About

Rokib Al Dhin Raadh, 18: founder of OXIEDO and inventor of ORMAS.

Founder and CEO of OXIEDO and inventor of ORMAS, a self-repairing neural network. He taught himself machine learning and built both without a lab or a co-founder. In 2026 he earned ten independent selections, including a DeepMath 2026 acceptance after double-blind review and a NeurIPS 2026 reviewer role. The full recognition record, a six-year timeline, his personal statement, education and contact follow.

Now

Recognition · 2026

Ten independent selections in 2026, at 18.

Rokib Al Dhin Raadh earned eleven independent selections in 2026, all from cold applications, across peer review, AI conferences and venture. In research: DeepMath 2026 accepted his stability result after double-blind review, NeurIPS 2026 made him a reviewer for its Trustworthy AI for Good workshop, Cosmos Institute ranked the work highest in its 2026 review cycle, IARCO 2026 named him a finalist, and Cohere Labs admitted him to its Open Science Community. In venture: Freshmango and 1752vc Ignite (top 1%) offered OXIEDO places, Onstage invited him to its W26 pre-pitch event, Entrepreneur First interviewed him in London, The Bridge in San Francisco took him to its final round, and 500 Global invited him to interview on 22 October.

Latest decision

Accepted30 Sep 2026

Stability result accepted at DeepMath 2026.

DeepMath is the Conference on the Mathematical Theory of Deep Neural Networks, this year hosted by Ohio State University in Columbus on 29–30 October, with an organising committee drawn from Johns Hopkins, Michigan and Ohio State. It takes theory only, and reviews it double-blind, so reviewers assessed the work without knowledge of his background. The paper, “Self-Repair as a Bounded Disturbance: Input-to-State Stability of Neural Network Training Dynamics”, was accepted for poster presentation. It formalises the stability guarantee underlying ORMAS. The 2026 invited speakers come from Stanford, Michigan, UPenn and the University of Washington, and past editions have been supported by the National Science Foundation and the Simons Foundation.

Two more are in motion: Onstage is ranking its W26 cohort by interest from 350 venture funds, and 500 Global interviews him on 22 October.

DeepMath 2026Ohio State University

Onstage

30 SepColumbus, OhioDeepMath 2026 · Mathematical Theory of Deep Neural NetworksAccept (Poster) after double-blind review. The input-to-state stability result behind ORMAS, presented at Ohio State on 29–30 October.Accepted

W26 cohortCentral LondonOnstage · W26 Demo DayInvited to the pre-pitch event in Central London. Onstage ranks applicants by interest from 350 VC partners, including a16z, Sequoia, Balderton and LocalGlobe.Invited

Since SepSan FranciscoEntrepreneur First · The BridgeTwo interview rounds completed; reached the final round. Details below.Final round

Status as of 30 September 2026

2026, month by month

Jul 2026

Cosmos Institute

Ranked highest in cycle

Aug 2026

Entrepreneur First

Cold application → first-round call

Sep 2026

DeepMath · Freshmango · 1752vc · IARCO · Cohere Labs

Peer-reviewed acceptance, two accelerator places, a research-competition final, Cohere Labs membership.

Oct 2026

DeepMath 2026

Poster, Ohio State, Columbus

Dec 2026

NeurIPS · AI4GOOD

Reviewing, Paris

The selections

Dec 2026

NeurIPS 2026 · AI4GOOD Workshop · Reviewer

Reviewer, NeurIPS 2026 Trustworthy AI for Good Workshop

Raadh is a reviewer for Trustworthy AI for Good at NeurIPS 2026 in Paris, reviewing submitted papers on mechanistic interpretability, attribution, auditing and post-deployment monitoring. The ICML edition of this workshop took 539 submissions and accepted 34%, assessed by a committee of 237 reviewers drawn from Oxford, MIT, Toronto, Berkeley, Stanford and Mila.

237Reviewers on the committee

34%Acceptance rate, ICML edition

Jul 2026

Cosmos Institute · Review Cycle

Ranked Highest in the Cosmos Institute Review Cycle

Cosmos Institute ranked the ORMAS application highest of every application in its cycle and invited Raadh to its technical track.

HighestRank in the cycle

InvitedTechnical track

Sep 2026

Freshmango · Equity-Free Accelerator · Accepted

Accepted to Freshmango, the Equity-Free Accelerator

On 22 September Freshmango offered OXIEDO a place in its equity-free programme after a single interview: access to a 5,000-founder network, $4M in AI credits and introductions across 160 venture funds, with no equity or fees.

0%Equity taken

$4MIn AI credits

160Venture funds

Sep 2026

1752vc · Ignite · Accepted

Accepted to 1752vc Ignite from the Top 1% of Applicants

1752vc placed the application in the top 5% of its intake, then held a final round to select the top 1%. OXIEDO was offered a place in the limited Ignite cohort, chosen from thousands of applications across three stages.

Top 1%Final-round selection

AcceptedIgnite cohort

Sep 2026

IARCO 2026 · International Academic Research Competition · Final Round

Finalist, IARCO 2026 International Academic Research Competition

The International Academic Research Competition is hosted by YRJ, sponsored by SaveMyExams with Domain.ME, and this year drew nearly 500 participants from 60 countries. ORMAS was among the research papers selected for the final stage, a recorded research presentation, and Raadh is listed as selected on IARCO's published 2026 finalist list.

~500Participants

60Countries

SelectedFinal stage, Junior

Sep 2026

Cohere Labs · Open Science Community · Member

Member, Cohere Labs Open Science Community

Cohere Labs is the research lab of Cohere, one of the leading foundation-model companies. On 25 September it welcomed Raadh into its Open Science Community, noting that the solo development of ORMAS across 383 controlled experiments “demonstrates remarkable initiative in ML safety and auditability.” The lab pointed to its BIRDS and Safety & Alignment programmes as a route to extending the architecture to a transformer backbone.

MemberOpen Science Community

ORMASCited in the welcome

Entrepreneur First · London to San Francisco

Aug 2026

Entrepreneur First · London · First Screen

Entrepreneur First: First-Round Interview, London

ApplicationSubmitted cold

Gate 1

ResponseFirst-round call booked

An unsolicited application to one of European venture's most selective sourcing programmes led EF's talent team to invite Raadh to a first-round interview.

EF's first screen assesses how a founder thinks rather than the idea itself. The conversation covered why interpretability is an architecture problem rather than a tooling one, and where ORMAS goes next.

It led to The Bridge.

ColdUnsolicited application

15 minFirst-round interview

Entrepreneur FirstThe Bridge — San FranciscoAug–Sep 2026 · Final Round

The Bridge is Entrepreneur First's eight-week residency in San Francisco, where they place a small cohort of technical founders alongside the US investor and operator network, and back those who come through it with company-building capital. It is EF's route for founders who have already built something.

Raadh completed two interview rounds and reached the final round.

Interview 1 · DoneInterview 2 · DoneFinal Round · Reached

What it adds up to

What the 2026 record establishes.

Peer review. DeepMath 2026 accepted the ORMAS stability result after double-blind review by specialists in deep-learning theory.

Standing in the field. NeurIPS 2026 brought Raadh in as a reviewer for the Trustworthy AI for Good workshop, reviewing papers in interpretability and model auditing, the area his own research addresses.

Independent assessment. Cosmos Institute ranked the work highest in its cycle. Freshmango and 1752vc each offered places after competitive selection, The Bridge in San Francisco took him to its final round, and 500 Global has invited him to interview. Every one began as an unsolicited application.

Made to be checkedEach item above has a named source and a date: the Cosmos correspondence, the NeurIPS reviewer assignment, the EF stage records, the Companies House filings. Documentation is available on request, and every claim is listed with its source on the Evidence page.Request documentation

Profile

Founder first: from a first venture at 12 to OXIEDO at 18

Founder Introduction · 31 sec

Five ventures, a new neural network architecture, and the company built on it.

Rokib Al Dhin Raadh, 18, is the founder and CEO of OXIEDO and the inventor of ORMAS, a self-repairing neural network. Raised in Dhaka, he ran five ventures between the ages of 12 and 17 and sold a stock-prediction system for about $10,000 at 15. Tracing one structural cause across those ventures led him to multi-agent systems (OXIMO), then to a new neural network architecture (ORMAS), and then to a company built on it (OXIEDO).

Two of the projects below are active, two are concluded and one is in specification.

In the video: who ORMAS is for, the gap it fills, and how it works.

Timeline

2020

Self-taught from age 12

Self-taught from the start, building each project a step beyond his current skills: a habit formed before any formal computer science.

12–17

2020–25

Five ventures between 12 and 17

Software services, cold-call automation, digital marketing, e-commerce and automation tooling. Each failed at a different layer, and together they pointed to one finding: the constraint was coordination, not individual capability.

2023

Sold a stock-prediction system for about $10,000

Built and sold at 15. The proceeds funded the RTX 3090 that later ran all 383 ORMAS experiments.

16–17

2023–25

OXIMO: a structural answer to coordination

A multi-agent system that decomposes work and creates specialist roles on demand. Rebuilt from a 106,000-line monolith to 40,933 lines with 2,069 passing tests. Now concluded.

2025–26

Black Bloxie LTD: a twelve-month field study

A UK company operated as a controlled test of OXIMO: the system was removed and restored to measure its causal effect (−91%, then +1,300%). The study also exposed the limit that led to ORMAS: the underlying models could not learn from noisy production data.

17–18

2025–26

ORMAS: 383 experiments on one GPU

A network that diagnoses and repairs its own damage during training, tested across four architectures and six noise regimes: 80.3% recovery where parameter-matched baselines remain at 10.0%.

2026

ORMAS breakthrough · DeepMath 2026 · NeurIPS reviewer

OXIEDO, founded in 2023, gets its core technology: ORMAS, licensed to five regulated sectors. The stability result is accepted at DeepMath 2026 after double-blind review, and Raadh becomes a reviewer for the NeurIPS 2026 Trustworthy AI for Good workshop.

Venture · failed Built Verified externally

Current Status

ORMASCurrent work

The research

Identifies, names and repairs its own failing components during training.

80.3%

recovery · baseline 10.0%

Open

OXIEDOLive

The business

Licensing the architecture to institutions that cannot yet train on their own regulated data.

Live

oxiedo.com · five sectors

Open

OXIMOClosed

Prior work

Turns a one-sentence brief into an organisation, creating the roles it lacks.

40,933

lines · 2,069 passing tests

Open

Black BloxieClosed

The experiment on OXIMO

A live UK company run for twelve months to measure the system’s causal effect.

−91% / +1,300%

12-month lesion study

Open

Project CherryNot built

Planned

A network that grows its own components. Fully specified; next to be built.

—

specification stage

Open

The ten 2026 selections are in Recognition. The five ventures are covered in Ventures, and credentials in Education.

Where to Start

Suggested starting points

Researchers

The ISS derivation, all 383 runs, the telemetry design and the adverse results. Scrutiny of the stability bound is especially welcome.

The research

Investors

The product, the regulatory drivers and timeline, and a full risk register.

The company

Engineers

40,933 lines and 2,069 tests: a system that turns one sentence into a working organisation.

The codebase

Sceptics

A twelve-month controlled study on a live company with real customers, with every figure disclosed.

The experiment

About

Personal Statement

Personal statement: from a first venture at 12 to DeepMath 2026 at 18.

In his own words, Rokib Al Dhin Raadh sets out how he built ORMAS and OXIEDO alone and self-funded, and the three resources the next phase requires: multi-node compute, capacity to train Project Cherry, and a control theorist.

Rokib Al Dhin Raadh · Age 18 · Founder & CEO, OXIEDO

14–16 hrs

A normal working day

Working days livestreamed

Self-funded

Every experiment on his own hardware

Solo

Sole author of ORMAS

100%

IP ownership

Part One

Built Alone, Self-Funded and on the Record

I started my first venture at 12 and ran five by 17. Each failed at a different layer, and together they showed me the constraint was coordination, not individual capability. At 15 I built and sold a stock-prediction system for about $10,000; that money bought the RTX 3090 that later ran every ORMAS experiment. OXIMO was my answer to coordination. Running it on a live UK company for twelve months showed that the models underneath could not learn from noisy production data, and ORMAS was my answer to that. OXIEDO is the company that licenses it. At 18, its stability result was accepted at DeepMath 2026.

I have built this work independently, without a university, research group, advisor or external funding. Every experiment ran on hardware I paid for, and every architectural decision was my own.

I work fourteen to sixteen hours a day. In November 2024 I livestreamed every working session for two weeks, producing 28 sessions of unedited footage, submitted as a Guinness World Record application. The full archive is linked below.

Work Archive — November 2024 · 28 Sessions

Unlisted livestreams recorded for the Guinness World Record application: complete, unedited working sessions.

01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28

Machine Learning

PyTorchCNN / ResNet / DAGTransformersPCGrad Gradient SurgeryConvergence TheoryNoise-Robust Training

Systems Engineering

Multi-Agent OrchestrationConstructor InjectionAtomic Rollback3-Tier MemoryAdversarial FilteringGraph-of-Thoughts Planning

Infrastructure & Web

DockerFastAPIPrometheusSSE StreamingHMAC WebhooksCI/CDPydanticSQLAlchemy AsyncChromaDBReact / TS / Web3

Programming Languages

Python (Advanced / Production)C++LaTeX

Spoken Languages

English (Fluent)Bengali (Native)

Part Two

Three Resources the Next Phase Requires

The foundations are in place: the code runs, the experiments are complete, the architecture is public and its stability result has passed peer review.

Working alone does not scale. What the next phase needs is not drive but calibration: colleagues, advisors and reviewers who shorten iteration cycles and catch errors early.

What the company does next is set out at oxiedo.com. The table below lists the three external resources that plan depends on.

Resource	Why it matters
Multi-node compute	The ORMAS protocol is designed to transfer to Transformers unchanged, as the mathematics is architecture-agnostic. Demonstrating that at scale requires multi-node H100/A100 access.
Capacity to train Cherry	Pre-training a language model on the three-signal architecture builds structural correction into pre-training rather than adding it afterwards through RLHF. It is a compute-intensive programme beyond a single GPU.
A control theorist	The ISS characterization passed double-blind review at DeepMath 2026 by deep-learning theorists. The next step is scrutiny from control theory.

Transformer-scale validation, Project Cherry and control-theory review of the ISS result all require an institutional environment I have so far worked without. That is why I am raising capital: to build a team, so the work no longer depends on one person to check it.

Rokib Al Dhin Raadh — 2026

About

Education

Self-directed by design: seven MIT programmes and a DeepMath 2026 acceptance before leaving school.

Rokib Al Dhin Raadh, 18, is in his final year of secondary school and chose to learn his field independently, without a university or advisor. He completed seven MIT Open Learning (MITx) programmes and Andrew Ng's Deep Learning Specialization, and learned whatever each problem required: control theory for the ORMAS stability proof, market microstructure and UK company law alongside machine learning.

12th Grade (Final Year) · Self-directed

Self-taught in machine learning, systems engineering and theoretical computer science, with a focus on underlying mechanisms. Each subject was learned because the work demanded it, including four well beyond software:

Control theory, for the stability proof

The ORMAS stability characterization rests on control theory, which required a working intuition for why a perturbed system settles, or fails to.

Where trust in automation ends

Automating European cold calls showed precisely where human trust in an autonomous agent ends: at the transaction. That boundary became a design constraint in his later systems.

Market microstructure, at fifteen

The stock-prediction system he built and sold at 15 had to account for institutional order flow, which cannot be inferred from price data alone.

UK company law, without a lawyer

Incorporating Black Bloxie LTD at UK Companies House from overseas, without a lawyer or agent, meant mastering the filing regime, anti-money-laundering requirements and director obligations.

MIT Open Learning · MITx

Seven MIT programmes. Before finishing school.

While still in his final year of secondary school, Raadh completed seven programmes from MIT Open Learning, covering foundation models, AI in medicine, energy and transport, and AI entrepreneurship: the same regulated sectors OXIEDO now sells into. After an admissions call about his work, he was offered a 40% scholarship on a further MIT programme.

7MIT programmes completed

40%Scholarship offered on a further MIT programme

SponsoredInto MIT Open Learning's Universal AI programme by FIAP

MITx · Foundation modelsUniversal AI Foundational ModelsView record ↗MITx · MedicineAI & Precision MedicineView record ↗MITx · MedicineHolistic AI in MedicineView record ↗MITx · EnergyAI and Sustainability: EnergyView record ↗MITx · TransportAI and Sustainability: TransportationView record ↗MITx · TransportAI for Transportation: From Concepts to ImplementationView record ↗MITx · VentureAI & EntrepreneurshipView record ↗

MITx Online sometimes fails to load shared records; full records, certificates and grade reports are available on request.

DeepLearning.AI

Andrew Ng's Deep Learning Specialization, the five-course programme from the co-founder of Google Brain and Coursera.

20+ further certifications

Machine learning, systems architecture and applied AI from Google, UC Davis and the University of Michigan. Verifiable on request from raadh@oxiedo.com.

FIAP · São Paulo

One of Brazil's leading technology schools made him a member in early 2026 and sponsored his place in MIT Open Learning's Universal AI programme.

Research over exam preparation, by choice

A deliberate choice to prioritise research.

Work on the ISS characterization and OXIMO took priority over exam preparation. The results are recorded not in a transcript but in 383 experiments, 40,933 lines of production code, a public preprint and a peer-reviewed DeepMath 2026 acceptance.

383

Controlled experiments

40,933

Lines of production code

DeepMath 2026

Accepted · Poster

UK Ltd

Incorporated, England & Wales

About

Contact

Review the research, reproduce the results, or discuss OXIEDO.

Rokib Al Dhin Raadh, founder and CEO of OXIEDO, takes direct enquiries from researchers reviewing the ORMAS stability result, investors and partners in compute and regulated data, and engineers reproducing the 383 experiments. Each has a starting point below.

Researchers

Stress-test the DeepMath 2026 stability result

The preprint is on Zenodo, and the codebase reproduces the core claims in under an hour on a single GPU. Feedback on the ISS stability result, including potential errors or tighter bounds, is particularly welcome.

Read the Preprint

Investors and partners

Fund the next phase: compute and regulated data

The next phase requires multi-node compute for Transformer-scale validation and a data partner with regulated data. The round, milestones and risk register are set out on the company site.

The Investment Case

Engineers

Reproduce the core claims in under an hour

A single script, reproduce.sh, runs every experiment in the paper, so every result can be independently verified or extended.

Explore the Codebase

Write to the founder directlyraadh@oxiedo.comReplies are typically sent within one working day.oxiedo.com