ODSA · A general optimization framework for interacting systems
Physics-grade answers, at the cost of arithmetic.
Supercomputing-class physical prediction on consumer hardware — auditable end-to-end, deployable at the edge, aligned with the EU AI Act by design.
01The assumption every industry inherited
Fast. Small. Certifiable. Pick two.
Every category of modern computing — green AI, IoT, wearables, medical devices, industrial control, mission-critical systems — exists as a way of living with one belief: you cannot have speed, footprint, and auditability at the same time.
Entire industries are organized around which two of the three they chose. Compute-heavy sectors rent clusters. Edge sectors accept weak models. Regulated sectors accept slow, certifiable pipelines.
A technology that removes the tradeoff does not fit any of the boxes — it explains why the boxes exist.
02The counterevidence
Numbers first, promises later.
ODSA is not a product inside a category — it is a layer underneath the categories: the same engine, wherever a system of interacting parts must be evaluated quickly and defended afterwards.
Ground state of a 1,000-qubit Heisenberg system, computed on a laptop CPU. The exact answer would require a state vector larger than the observable universe.
Error against exact methods at every scale where they remain feasible — verified, not asserted.
Cheaper than a conventional DFT screening campaign — €10k–100k and 4–40 days become one hour at ~€0.
Predictions per second on plain CPU — no GPU, no cloud, zero marginal cost per answer.
Source: Ciber Fabbrica validation runs & technical reports, 2026 — available on request
03How that is possible
Small is the feature, not the limit.
Physics-informed features
Every input is a named, bounded physical quantity with a documented derivation. Nothing learned, nothing hidden.
A small transfer function
3.8k–165k parameters — it behaves like a formula that can be characterised by ablation, not a black box.
Prediction + audit trail
Deterministic output that decomposes back to its contributing physical quantities. Inspectable by a domain expert.
Small is what makes it auditable — and auditable is what makes it deployable where certification decides what ships.
04ODSA Q · Native quantization
A complete predictive solver in 1KB.
Trained and quantized natively — not shrunk after the fact. For twenty years, edge AI negotiated a three-way trade between accuracy, footprint and autonomy: pick two. The negotiation is over.
A complete predictive solver — small enough to fit in the spare memory of almost any microcontroller ever made.
Accuracy retained at sub-kilobyte on NASA C-MAPSS engine prognostics — score 3,214 vs 3,028 at full precision.
Retained on 3D quantum many-body ground states (error 3.16% vs 3.08%) — the method carried over from the other domain unchanged.
Smaller footprint than typical published solutions for the same industrial benchmark.
The controlConventional after-the-fact compression destroys these same models — errors thousands of times worse. The margin belongs to the method, not to the trick.
Honest positioningSub-kilobyte retention is measured today in two deliberately unrelated domains — portability evidence is strong, but it is two points, not twenty. Each new domain is validated before it is claimed.
05Why sub-kilobyte changes the math
Intelligence that stays on the device.
On battery devices, the radio is the energy budget. Not the processor — the radio. Every kilobyte sent to the cloud costs battery, latency, privacy, and a connectivity subscription. Intelligence that lives on the device is nearly free to operate; intelligence that phones home is a permanent running cost.
It fits the memory devices already have
Sub-kilobyte models live in spare flash and RAM — alongside existing firmware.
So it ships as a software update
No new hardware, no recall, no redesign cycle.
To devices never designed for intelligence
Sensors, wearables and controllers already in the field gain real prediction.
The installed base becomes the market — retrofittable intelligence, delivered over the air.
06The categories, reorganized
One engine, one map.
Every category you know is a point on this map. Two properties decide who can play in the top-right: the decision must happen at the data, and it must survive an audit. ODSA holds both, by construction.
07Methodology
Trust is built in steps.
Exact ground truth
Small systems solved exactly. Error measured, not estimated: 0.1–0.8%.
Public benchmarks
NASA C-MAPSS, PEPBI, 130 public battery cells — ground truth anyone can re-check.
Literature regimes
QMC and DFT references where published values exist — agreement documented per regime.
Documented extrapolation
Beyond exact reach, methodology and code are open for review — flagged as extrapolation, never sold as proof.
We claim certainty only where it can be checked — and we show the ladder openly above that.
08Defensibility
Why this position holds.
A proprietary core
The feature construction is the moat. Internals stay undisclosed while every result stays open to independent verification — protection without opacity toward the customer.
A regulatory head start
Deterministic, traceable, bounded by design — the properties the EU AI Act demands of high-risk systems are properties ODSA already has. Competitors face a retrofit; we face a paperwork exercise.
Compounding validations
The architecture does not change between domains, so every new validated sector de-risks all the others. Six audited domains today; each evaluation engagement adds another public proof point.
09The commercial shape
Evaluation first. Embedding later.
Send one problem
Inputs and ground truth — or your current reference method. Under NDA.
We run it
On our own laptop — no access to your infrastructure. Hours to days, not months.
You compare
Accuracy, speed, cost — against the method you use today, on your data.
Let the numbers talk.
If the numbers hold, we talk licence. No fee for the first evaluation.
You don't buy ODSA; your product gets ODSA inside — inside the dispatch controller, the screening pipeline, the design tool, the monitoring stack. Deployment measured in hours, no infrastructure to add.
10The ask
One problem. One comparison. One honest answer.
Design-partner pilots with organisations that have hard physics bottlenecks — and an open investment conversation. If evaluating one candidate cost nothing and took a millisecond, what would you do differently?
Technical reports on request · The construction is proprietary, the evaluation is open