01
A licensing company. What it licenses is train-time accountability: the architecture that makes a neural network keep a bounded, logged account of every change it makes to itself, as it makes it, which is the account that says what the model took from each source it was given. It exists in no standard architecture and no tool recovers it afterwards. It has to be built into how the model learns. We built it, and we license it to institutions holding data under obligation.
02
Data is the asset of the next decade and most of the valuable part of it is locked. Hospital records, bank ledgers, assay runs and licensed corpora sit behind regulation and contract, unused, because handing them to a model means handing them to something that afterwards cannot say what it did with them. Enough of it to change how medicine is practised, how risk is priced and how systems are built to fail safely, and almost none of it usable. The constraint is an accounting one rather than a technical one. Nobody holding that data is being careless, and nobody holding it is wrong to hold it under the conditions that exist. What is missing is the account that would make opening it defensible, and that account cannot be recovered later. It has to be built into how the model learns.
03
Turning one property into one product, licensed to institutions that hold data under obligation. Transparency is the mechanism rather than the goal. The goal is that data currently locked by regulation and contract becomes data an institution can defensibly train on, because the record now exists to show what was taken from it and to take it back out on request. Behind those doors are applications nobody has reached, in sectors that have never had a model trained on their own data because the data was never releasable. The mechanism is published so it can be checked, and the release is committed in the licence rather than promised in a blog post.
Rokib Al Dhin Raadh founded Oxiedo and leads it as CEO. He is the inventor of ORMAS, the training architecture the company licenses: a neural network that records every change it makes to itself while it trains, names the part that broke, and repairs it.
He started on the question in 2023: can a neural network give a truthful account of what it does to itself, while it does it? He spent 2024 and 2025 on research, the first attempt failing against its own destruction suite, and reached the first working version on 11 July 2026. He designed the architecture, wrote the code, ran every experiment and wrote the paper, which has been public since 1 August 2026. On 30 September 2026 his paper on the stability theory behind it was accepted for a poster at DeepMath 2026, the Conference on the Mathematical Theory of Deep Neural Networks.
Before ORMAS he ran a twelve-month AI-run e-commerce experiment. He is based in Dhaka, and he is the person who signs the licence and answers for every claim on this site.
Nothing here was planned from the beginning. The work began as a research question two years before there was a legal person behind it, the company went where the problem was, and the problem turned out to be considerably larger than the one it set out to solve.
Oxiedo exists from 2023, quietly and without much happening in it. What it had was a question nobody had a good answer to: can a neural network produce a truthful account of what it does to itself, while it does it, rather than a reconstruction assembled afterwards and hoped to be accurate? Everything since has been the attempt to produce one somebody else could check.
The field had priced opacity as the cost of capability and moved on. Looked at closely, that trade was not a law of the method. It was the consequence of one early decision, reasonable when it was made and never revisited since. Acting on that took several attempts, and the early ones did not work. An architecture was built, put through a destruction suite written in advance, and lost to a plain CNN at every noise level tested. The correction mechanism never triggered once across 42 runs. Those logs were kept rather than deleted, and they are on this page further down.
The research reaches the point of needing a legal person behind it, something that can hold an agreement, take a payment and issue a licence, so it gets one: contracting runs through an existing UK-registered company, named in full in any agreement, while the permanent structure is decided. The work was then put in front of machine-learning practitioners and people working inside the regulated sectors, for feedback on what a custodian would need before releasing any data: a calibrated bound, a record that survives an audit, a named counterparty who will stand behind both. That feedback shaped the architecture more than any benchmark did. None of it was a customer or buyer conversation; those have not begun. No custodian has yet released data to us, and that remains the largest single thing standing between the result and the field it was built for.
The trainer rewritten and the architecture reconstructed, the suite that destroyed the previous version is cleared for the first time. Not a refinement of the earlier attempts. A different construction, arrived at after the earlier ones had been taken as far as they went and abandoned.
The complete system: 383 controlled experiments across four architecture families, 67 archived runs reproducible from seed, and 16,316 lines of instrumentation. Every result on this site was measured on or after this date. It is worth being exact about the timing, because it explains the rest of the position. Before that date there was nothing here an institution could have been sold, and selling it anyway would have broken the operating principle below that says nothing ships while a known defect is open.
The stability analysis behind ORMAS, “Self-Repair as a Bounded Disturbance: Input-to-State Stability of Neural Network Training Dynamics”, is accepted for a poster at DeepMath 2026, the Conference on the Mathematical Theory of Deep Neural Networks, held in Columbus, Ohio on 29–30 October; it is presented on 30 October. It is the first time the work has been reviewed and accepted outside the company.
The architecture, the results and the reasoning are released in full on publication, and what is sold is not the research but a licence to the accountability it produces. Commercial licensing is open, pre-bookings are being taken, and we are open to early investment conversations. The first dataset partnership is the priority ahead of everything else, because a custodian releasing regulated data to this architecture is the proof the whole thesis rests on. The Delaware C-corporation is formed when the pre-seed closes. The founder is based in Dhaka, and the company relocates when the first design partner gives it a reason to.
Rokib Al Dhin Raadh
Founder & CEO
A custodian deciding whether to release regulated data asks for three things before anything else: a calibrated bound, a record that survives an audit, and a named counterparty who will stand behind both.
The first two are architecture. The third is not something an architecture can supply. A published mechanism has no counterparty of its own, and a bound with nobody behind it cannot be entered into a filing or defended in a review. What a licence adds to a published mechanism is a person who signs the declaration and answers for it when a regulator, an auditor or an opposing expert puts it under pressure.
Rokib Al Dhin Raadh is that person here. The architecture, the results published against it and every run where it loses are attributable to one named person, the founder and CEO, and every archived run behind them regenerates from seed, so the claim can be checked rather than taken on trust. Corrections are issued under the same name as the original claim.
That is a deliberate structure rather than a stage the company is passing through. Accountability distributed across a vendor is accountability nobody can locate, which is the reason a compliance function asks who signs before it asks how the method works. A licence that function can rely on names someone, and this one does.
Each of these costs something. They are listed because a company is what it does when the decision is expensive, and because every one of them is checkable against the rest of this site.
The architecture and every experiment behind it are released on publication and free for research from that day. The intention is to contribute a fact to this field rather than an opinion: measured, reproducible, and checkable by anyone who cares to check it, which is why every experiment regenerates from seed. Credibility here rests on results anyone can reproduce, and a permanently closed method is worth less to a validator, not more. If the result is wrong, the field establishing that quickly is the better outcome.
The read-only mode is not a stripped tier, it is the default posture. Nobody sensible lets an unproven system modify weights inside a run that costs six figures, and the product does not ask them to.
The modification record currently exists in three incompatible shapes in three code paths, and one of them silently drops compliance fields. A compliance artefact assembled from three incompatible records is not a compliance artefact, and that closes before anything is sold on it.
The study that decides four of the seven features runs a third of the way through the build, not at the end. Running it late costs the same and buys nothing. Running it early preserves the option to change course.
Every run where it loses is published on the technology page: adversarial weight injection (−1.0 pp), a 100× weight explosion (−0.9 pp), and ResNet-18, where an uninstrumented network recovers slightly better (92.3% vs 91.6%). In fields whose central complaint is that vendors overstate their approximations, being exact about their own limit is the position rather than a caveat on it.
A bank, a hospital and an AI lab buy the same mechanism and call it by three different nouns, because the thing all three are really buying is the ability to use data they already hold. That is a property of the architecture rather than a marketing decision, and it is why there is one licence rather than seven contracts.
An earlier version of this architecture was put through a destruction suite built for the purpose: 42 runs, two backbone types, and an explicit pass/fail list written in advance by the person who most wanted it to pass.
FAIL: Lost to a plain CNN at every noise level tested.
FAIL: Under weight-perturbation shock, the plain CNN recovered better.
FAIL: The correction mechanism never triggered once across 42 runs.
FAIL: The ablation arms were indistinguishable from the full system.The logs were kept rather than deleted. The trainer was rewritten, the architecture reconstructed, and the suite that broke the first version became the standard the second had to clear.
The evidence base is adversarial, not confirmatory. That is the answer to anyone asking whether the results were selected after the fact.
A dataset partner above all, then a commercial lead, a research engineer, and two single consultations that between them unblock the largest part of the product. The commitment shape is stated on each one, because some of them are not jobs.
Entity
Oxiedo · A licensing company · Delaware C-corporation formed at close · Founder based in Dhaka · Until then, contracting and payments run through an existing UK-registered company, which is named in full in any agreement
Contact
Every message is read by a person and answered within a few days. We are open to early investment conversations ahead of a priced round.
Get in touch