Small changes. Vast outcomes.
We use data to expose the hidden danger in the local environment and surroundings.
Investigative agents built on small language models — 0.5–8B, mission-trained, air-gappable — that find the infinitesimal change before it becomes the irreversible one.
Every system gives a warning before it commits. The warning is small, it is early, and it looks like nothing. Clustral is built to catch it.
Two runs of the same system.One started 0.00001 apart.
The Lorenz attractor — where the butterfly effect was found, in 1963.
Cheap enough to observe every event — not a sampled few.
Each flag becomes an evidence-backed case, not another alert.
While the change is still small enough to be worth deciding about.
Built for regulated, board-supervised environments
A system does not fail all at once. It drifts — quietly, cheaply, reversibly — and then it crosses a line, after which every remaining option is worse than the one you no longer have.
Illustrative: a contact report delayed twenty minutes by a degraded uplink — the window closes, the option is gone. The uplink degraded hours earlier. Nothing alerted, because nothing had broken yet.
By definition. If it were large, you would already have seen it. Detection has to be cheap enough to run on everything, continuously — which rules out sending every event to a large model.
The distance between the first observable signal and the point of no easy return is measured in minutes, not review cycles. An alert that queues for a human has already spent the budget.
The asymmetry is the whole argument. The cost of investigating a thousand small things is a rounding error against the cost of one irreversible one.
Two identical machines, installed to the same spec. One fails twice a year; the other never has. Nothing is wrong with either machine — the difference is entirely where they stand. Everyone instruments the asset. Almost no one investigates the conditions around it.
Weather, terrain, air, water, power, the plant upwind, the road that dusts the intake, the frost pocket nobody mapped. These conditions are public, continuous and largely unread — which is exactly why the danger hides there.
Same asset. Same spec. The difference is entirely in what surrounds it.
Hyperlocal observations, storm and radar history, climate normals — what actually passed over this location, hour by hour, not the regional average.
Elevation, slope, drainage, land cover and exposure. The reason one site floods, corrodes, ices or overheats and its twin does not.
What operates nearby and what it releases — emissions, dust, vibration, discharge, interference — and whether the wind carried it here.
Where a thing is, when the conditions converged, and who else was affected in the same window. Coincidence becomes correlation becomes cause.
Most environmental tooling ends at a colour and a number — amber, 0.72 — which is precisely what fails the moment it matters, because a score cannot be audited, contested or relied upon by anyone who carries liability. Our agents finish somewhere else: a case file.
Because the models are small, this runs where the data already lives — on-premises, at the edge, or fully air-gapped. Sensitive site data never has to leave the building to be investigated.
The FAA publishes where GPS may fail. Nobody measures whether it did. Aeroscan samples live aircraft across and beyond each published test footprint — Memphis, Louisville, Anchorage, White Sands, the Nevada ranges — and reports what aircraft actually experienced.
It is the thesis in one screen. Not a score — a case file: the finding, the sample it rests on, the aircraft showing degraded integrity outside the published boundary, and a plain statement of what the evidence cannot yet support.
Open AeroscanBuilt on open ADS-B and published FAA advisories. Indicative of interference conditions — not a navigation aid, and not for operational flight use.
Touch a hot stove and your hand moves before your brain understands what happened. The spinal cord reacts; the cortex is informed afterwards. Clustral is built the same way — the cheapest reliable computation runs first, and escalation is earned.
The biggest misconception in AI is that every piece of data should meet one large model. Intelligence is layered, as it is in nature. Every higher layer is more capable, more expensive — and therefore sees fewer events than the one beneath it.
No context, no understanding — they just notice deviation from normal, in microseconds on almost no power. Cheap enough to watch every event, forever.
Statistics see snapshots; these watch movies. Is reality itself shifting, and will this trend cross a dangerous threshold?
Classify the anomaly, trace the causal chain, assemble evidence, rank hypotheses, write the brief. Sherlock Holmes, not Wikipedia — the output is a case, not a score.
Only the rarest events reach a person — and they arrive with the evidence already assembled, not as an alert to go and investigate.
They synthesise reports, generate training data and improve the next generation of investigators — but never make a live call. That is what keeps the system deterministic, auditable, and able to run air-gapped.
Funnel volumes are illustrative, chosen to show the shape of the economics rather than any single deployment. Actual ratios are set by your event rate and risk appetite, and are agreed in the SOW.
Statistical reflexes keep rhythm. Temporal models preserve continuity. SLM investigators perform the solo. Frontier models compose the improvements that ship next quarter. No single instrument is asked to be the whole ensemble.
Every observation deserves a reflex. Only a few deserve investigation. Only the rarest deserve human attention.
An engagement is the ensemble, built against your event streams and your risk appetite. Everything below is in service of one outcome: a flag you can act on before the threshold.
Statistical and temporal layers wired into your streams — deviation caught at event rate, then persistence and trajectory tested on whatever survives that. Tuned to your baseline, so the reflexes fire on your normal rather than a generic one.
Mission-trained small models, fine-tuned on your domain and your incident history. They classify, trace the causal chain, rank hypotheses and write the brief — running on your hardware, air-gapped if that is what your perimeter demands.
The part that makes it defensible: evals on every model, lineage on every decision, append-only audit trail, ISO 42001 and EU AI Act evidence packs you can hand to a regulator.
Adjacent work — data platforms, integration into SAP / Salesforce / ServiceNow, enablement for your team — is scoped where an engagement needs it, not sold as a separate line.
You're likely choosing between a brand-name consultancy, a body-shop integrator, or a packaged SaaS. Each fails at a different point of the lifecycle. Clustral closes the seam.
Beautiful strategy, no production system. The recommendation is months old by the time engineering owns it.
The Clustral difference
Our strategists build with the engineers who will ship it — so what we advise is what we deploy.
An army of mid-level developers, no senior judgement on data, evals or governance. Cost balloons; risk hides.
The Clustral difference
Senior engineers stay on every engagement. Fixed value milestones, written scope, no resource-pyramid.
Black-box scoring, no audit trail, no path to your data realities. Procurement can’t sign; legal won’t clear.
The Clustral difference
Open architectures, citation-bound outputs, ISO 42001 / EU AI Act evidence packs you can hand to a regulator.
The components we ship inside engagements — eval harnesses, citation-bound generation, PII redaction, evidence vault — are also available as named modules with annual subscription pricing. Buy them à la carte, or get them embedded as part of an engagement.
Append-only audit-trail vault for every AI decision your enterprise makes. EU AI Act Art. 12 logs · ISO 42001 evidence pack · reproducible replay.
Continuous LLM & ML evaluation as a service. NIST AI RMF trustworthy-AI scorecards · fairness, robustness, hallucination · CI gates.
PII / PHI redaction at ingest, before any model call. GDPR + DPDP + HIPAA + LGPD lawful-basis aware · reversible for authorised reviewers.
Regulated-document intelligence: layout-aware extraction, citation-bound summaries, reviewer console with reject/correct retraining loop.
All four modules are currently in private beta. Indicative pricing assumes a single workload, single region; multi-tenant and on-prem deployments are scoped per engagement. Bundled with services engagements at a material discount.
Most discovery calls open with the same question — “is this a high-risk system, and where?” The Compass answers it across every jurisdiction at once. Pick a domain, an archetype, and an impact level — get a live global risk class with side-by-side treatment under EU AI Act, NIST AI RMF, ISO/IEC 42001, OECD, Singapore, China, plus privacy and sector overlays. The rule engine runs entirely in your browser. The optional AI mode sends your description to a hosted model — see Trust for the provider, region and retention.
Classify your AI use case across every major regulatory framework worldwide — EU AI Act, NIST AI RMF, ISO/IEC 42001, OECD, Singapore, China's GenAI Measures, plus privacy and sector overlays — and surface the controls a Clustral engagement would put in place.
Computed locally from a published rule matrix; not legal advice. Production engagements include a full conformity assessment with our counsel and your DPO.
Each of these is a live system, not a mockup. The pattern is the same every time: the signal existed, it was public or already collected, and nobody was reading it closely enough to act while acting was still cheap.
Aeroscan watches live aircraft behaviour across and beyond each published interference footprint and reports what aircraft actually experienced — by tail number, altitude and minute.
413 aircraft tracked over one freight hub · the single jet flying on degraded GPS, named
Live open aviation data. Indicative of interference conditions — not a navigation aid.
Lifecycle Hub scores every buried segment from what surrounds it — material, age, soil, pressure zone, break history — and ranks the handful about to fail, with the reasoning attached to each one.
1,892 segments · a risk threshold raised · the short list that survives it
Runs on synthetic data modelled on a real municipal network.
Ignition investigates the idea against the live web before anything is built — competitors, real customer complaints, name availability — and returns a case file with a verdict and every exhibit behind it.
an idea in · 21 exhibits gathered · a viability verdict you can argue with
Assess reads the occasion rather than the keywords, reduces the field to three defensible choices, and shows the reasoning for each — so the decision can be checked instead of guessed.
an occasion and a budget in · three ranked matches out, each with its reasoning
We are not the firm that arrives with a deck and disappears at deployment. Every system we ship carries its own evidence — the evals it passed, the data it touched, the policies it answers to. Senior engineers, written milestones, full handover.
Four commitments · written into every SOW
Asserted
The boards we work with don't ask “can AI do this?” — they ask “can we defend it when it does?” Every system Clustral ships is engineered to answer that question, in writing, before it goes live.
Every model we ship is paired with the policy, controls and evidence to operate it safely. We embed governance from day one, not as an afterthought.
Clustral AI Labs is a specialist AI engineering and governance firm. We pair senior strategists with hands-on ML, data and platform engineers, so what we recommend is what we can build. No layered handoffs, no middle-tier mark-ups, no AI theatre.
We deliver on cloud and on-prem, across overlapping APAC, EMEA and Americas hours, in any major currency. Every engagement comes with a written scope, fixed value milestones, and a governance package you can show your board, your regulator and your customer.
A 30-minute conversation is usually enough to know whether AI is the right lever — and where to pull it first. No deck. No pitch. Just engineering judgement.
NDA on request. Indicative response within one business day across APAC, EMEA and Americas hours.