Work
Founder of OXIEDO. Building companies since the age of 12.
Rokib Al Dhin Raadh is the founder and CEO of OXIEDO, which licenses ORMAS, his self-repairing neural network, to regulated industries. Each step built on the last: five ventures from age 12 and a ~$10,000 sale at 15; OXIMO, a 40,933-line multi-agent system with 2,069 tests; a 12-month controlled study on a live UK company, Black Bloxie LTD, where removing OXIMO cut output 91%; then ORMAS and OXIEDO.
OXIEDO lets regulated industries train AI on data they already own.
OXIEDO, founded by Raadh in 2023, licenses ORMAS on-premise to banks, hospitals and other regulated institutions. Their records are among the most valuable training data in existence, and most of it sits unused, because conventional models cannot account for what they learned from it. ORMAS writes that account during training: the Model Change Record, a tamper-evident log of every weight change in a run. It is built for SR 26-2 model-risk validation and for the EU AI Act's high-risk record-keeping obligations, which apply from December 2027.
Problem
Institutions do not buy audit trails for their own sake. A hospital wants to train on its own patient records without regulatory exposure; a bank wants a model its validators can sign off. Transparency is the mechanism; the value is access to data these institutions already own but cannot yet use. Interpretability tools sell an explanation after the fact; OXIEDO licenses the ability to train on that data in the first place.
ProductOn-premise by design: the data never leaves the building
The Intelligence Is Rented
So your data has to travel to reach it.
For a hospital, a fund or a defence programme, sending data off-site is often ruled out before accuracy is even considered. The framework sits inside the perimeter, but the intelligence does not.
The Training Happens Inside
So nothing has to cross the wall at all.
Everything stays inside the perimeter. For these buyers this is often the only deployable configuration, and it is why the record is produced during training rather than recovered afterwards.
MarketsFive regulated markets, one question regulators ask
PositionCurrent position: founder-led, licence defined
OXIMO: a 40,933-line multi-agent operating system that ran a live company for a year.
OXIMO is the multi-agent operating system Raadh built alone after five ventures showed him the same constraint: coordination, not individual capability. Given a one-sentence brief, it designs the organisation needed to deliver it and creates specialist roles when none exist. It ran Black Bloxie LTD, a live UK company, for twelve months. Its limit on real-world data is the problem ORMAS was built to solve.
Most multi-agent frameworks require a human in the loop to frame each task, prompt each model and pass outputs between steps. OXIMO's agents decompose work themselves, retain memory across sessions, create new roles and operate as an organisation.
Not a chatbot: an organisational structure that assembles itself.
Raadh rebuilt OXIMO from a 106,000-line monolith into 40,933 lines across 11 independently deployable repositories: 72% smaller, all 12 critical algorithms intact, all 2,069 tests passing. The architecture survived a complete reconstruction.
View the codebase · Production code, built to operate a live company.
Each OXIMO agent is built on a third-party model whose internals cannot be inspected. That was workable on clean inputs, and the system operated a live company for a year (see Black Bloxie).
Training on the company's own data changed that. Production data is mislabelled, contradictory and partly corrupted, and the established methods for handling it (DivideMix, ProMix, CoDE) all failed. The problem sat inside the model, beyond the reach of better orchestration.
OXIMO posed the question; ORMAS, the self-repairing neural network Raadh then invented, is his answer. It is now his full-time focus, and the technology OXIEDO licenses.
From one sentence to an organisation
Most agent frameworks require agents to be defined in advance. OXIMO derives the organisation from the brief and creates specialists that did not exist when the task began.
What Happens to a Task
A task arrives as one sentence. Nobody has said who should do it, how to split it, or whether the right specialist exists yet. An organisation assembles itself around the work and a validated deliverable comes out the other side.
- Task inone sentence
- Safety Gate7-check cascade
- Dynamic Routercomplexity + skills
- Executeone of four paths
- Sacred ChainMaster → Dept → Emp
- Validatemulti-stage merge
- Learnbrain updated
A1 · A2 · B1 · B2 — the B-paths hire a specialist mid-task
Creating New Roles Mid-Task
Most agent frameworks ask you to define your agents up front. When OXIMO meets work no existing role can handle, it designs the role, checks it is not a duplicate, validates it, tests that it produces coherent output, and commits it — with no human at any step. This is the orchestration-layer ancestor of Silent Node Injection.
Step 6 is why this is safe to run unattended: a partially-created agent is impossible. Either the whole hire commits, or none of it did.
Persistent, Compounding Memory
Three memory tiers, plus an Ebbinghaus decay curve so unused knowledge fades and reinforced knowledge strengthens. Agents mature Nascent → Learning → Mature → Expert.
This is what the ablation measured. Removing an embedded system does not migrate it — it discards everything the system had learned, and output fell 91%.
Cost Efficiency
OXIMO does not call one expensive frontier model. It cascades cheaper specialised models across six stages, each tuned to one kind of work — for equivalent output quality. Costs per task at model prices during OXIMO’s development (2024–2025).
A ~51× reduction, achieved by specialisation rather than compromise. It is also the argument for Cherry in miniature: many small specialists beat one large generalist, at a fraction of the cost.
Constructor injection everywhere, no global state, every dependency declared and every boundary enforced. Each repo can be tested, deployed and replaced on its own.
Removing OXIMO from a live UK company cut its output 91%.
Black Bloxie LTD, incorporated in England and Wales on 11 September 2025 (company no. 16711223), was Raadh's twelve-month controlled field study of OXIMO on a live e-commerce company, with a real supplier and customers in ten countries. When he removed the system, output fell 91%; when he restored it, output recovered and overshot its original level. Revenue alone cannot show whether the market would have delivered the same result. A removal test can.
Scope. This study tests OXIMO, an orchestration layer built on third-party models. It makes no claim about ORMAS, which is separate work with its own evidence.
The study in four phases
- Phase 1Aug–Oct 2025Cold startBaseline
- Phase 2Nov 2025–Feb 2026System removed−91%
- Phase 3Mar–Jul 2026System restored+1,300%
- Phase 413 Aug 2026High-value test$4,386
The complete revenue record for the twelve-month study.
A single order accounts for 65% of total revenue. It was the study's final test: whether an autonomous system could carry a customer through a high-value purchase. Total revenue is under $7,000, deliberately capped for the duration of the study.
The falsification ladder — seven questions, twelve months
Each rung is a test the thesis could have failed. They are in the order they were run, with the verdict on each.
- 01Phase 1 · Aug–Oct 2025 · Cold start
Can an autonomous system originate commercial demand at all?
The hardest available starting point: a brand-new store (incorporated as Black Bloxie LTD on 11 September 2025), zero domain authority, zero brand equity, zero advertising budget, no human in the acquisition loop. The store went live cold. Customers arrived, every one through LLM-referred channels.
Yes. Demand originated where there was none.
- 02Phase 1 · Baseline established
How much can it sell, and at what unit economics?
A capability claim needs a cost structure. The pipeline produced a complete 12-asset product suite for $0.0043 against a $50–$150 human benchmark. Baseline output was locked across all channels as the control condition.
The marginal cost of output collapses to near zero.
- 03Phase 2 · Nov 2025–Feb 2026 · The lesion
Strip the system out. Was it ever the architecture, or was it the market?
The central test of the study. Every generated asset was removed and the store reverted to bare platform defaults — same products, same prices, same supplier, same category, same seasonality. One variable withdrawn.
The architecture was the causal variable, not the market.
- 04Phase 2 · Dead zone
What does complete system removal look like?
Not decay. Cessation. Acquisition went to zero and stayed there for four months. The collapse landed across every channel on the same day rather than tapering channel by channel — the specific signature that rules out seasonality and macro conditions.
−100% acquisition. Simultaneous, not gradual.
- 05Phase 3 · Mar–Jul 2026 · Re-injection
Is the collapse reversible?
A one-directional result is only half a proof. If the system caused the output, restoring it should restore the output — and if it does not, the causal claim fails from the other side. Redeployed into the same entity with no other change.
Yes. The dead zone was a lesion, not a wound.
- 06Phase 3 · Recovery ablation
Does it return to baseline, or past it?
It overshot: the study’s most significant result. The knowledge graph had retained everything learned in Phase 1, so the system restarted with everything it had learned in Phase 1 (August–October 2025) instead of starting cold. It came back knowing things it did not know the first time.
Recovery compounds. Institutional memory survives.
- 07Phase 4 · 13 Aug 2026 · High-value test
Can it close a high-value transaction, or does trust break at the price point?
The residual objection: low-consideration purchases are easy. So the final phase listed high-value inventory and let the architecture run the entire funnel unassisted — discovery, positioning, objection handling, close.
A $4,386 order closed at $0.00 acquisition cost.
The study was capped deliberately. No jurisdiction has yet settled who is liable when an autonomous agent enters a contract, and a clean experiment was preferable to testing that boundary.
The technical reason was more important. The natural next step was to let the system learn from the company's own data, which, like all production data, was mislabelled, contradictory and partly corrupted. Every established method for handling it failed. That limitation is why Raadh's next project was ORMAS, a self-repairing neural network, rather than another product.
Five ventures from age 12, and a ~$10,000 sale at 15.
Raadh founded his first company at 12 and four more by 17, and at 15 he built and sold a stock-prediction system for about $10,000. Each venture hit a different limit. Together they identified the constraint every later system was built to remove: coordination, not individual capability.
A stock-prediction system, built and sold at fifteen
His first exit, at about $10,000. The proceeds funded the RTX 3090 that ran all 383 ORMAS experiments, so the research behind OXIEDO was financed by a product he built and sold himself.
Five ventures, five limits found
- 01Software ServicesAge 12–14Limit found: Distribution without credibility
The product could be built; reaching buyers was the obstacle. Enterprise procurement does not route on technical merit alone, and distribution without institutional credibility is not a technical problem.
- 02Cold-Call Automation, European MarketsAge 14–15Limit found: Trust limits at the close
Automation reliably handled 90% of the sales cycle, but every close required a human. Trust in autonomous agents stops at the transaction, a constraint the Black Bloxie study later tested directly with a $4,386 order closed with no human involved.
- 03US-Targeted Digital MarketingAge 15Limit found: The wrong bottleneck
The assumed bottleneck was distribution; the real one was production. The need was not to move content but to create it at volume.
- 04Organic E-CommerceAge 15–16Limit found: Coordination at scale
Optimising pricing, listings or marketing individually yields a few percent. The real challenge was coordinating research, listing, pricing, fulfilment and service at once, which no single operator can sustain.
- 05Automation ToolingAge 16–17Limit found: Confirmation of the pattern
The fifth venture confirmed the pattern: the constraint was coordination, not individual capability. That finding became the design brief for OXIMO.
Each venture failed at a different point, and all for the same reason. The constraint was coordination, not individual capability.
A structural problem needed a structural answer. That conclusion produced OXIMO; OXIMO's limit on real-world data produced ORMAS; ORMAS is the technology OXIEDO licenses.