Decision Lab · current proof and applied research
Run current decision tools, inspect Auralink and Reachy evidence, and see how product, architecture, model, safety and field assumptions are turned into decision records. The planned physical Foundry remains clearly labelled further down.
Digital lab live · physical Foundry planned
Runnable demonstrations, measured bench work and Hyperion-owned R&D are available now. The proposed physical facility, zones and missions remain illustrative until built, tested and labelled otherwise.
Available now
These are current digital, bench and internal-R&D artefacts—not the planned physical facility. Each page states what it can and cannot prove.
Flagship · deterministic · release 1.0
Configure a fictional AMR, drone, cobot or vehicle-feature product. Stress economics, reliability, capacity, operating envelope and delivery method; inject failures; then export a versioned Product Decision Record and Product Bridge Contract with explicit human authority.
Auralink
Pre-production · Hyperion-owned R&D. An inspectable reference implementation—not a client deployment or commercial-outcome claim.
Reachy Mini + SO-101
Measured · simulated · dry-run. Real sensing and teleoperation evidence stays separate from simulated autonomy and unexecuted robot actions.
Inspect the robotics evidenceMeasured on the bench
AMD Ryzen AI Max+ 395 (Strix Halo) · 128 GB LPDDR5X-8000 unified memory · medians of three runs after a discarded warm-up · no tuning
Read the complete field notes06 · Available now
These are current digital, bench and internal-R&D artefacts—not the planned physical facility. Each page states what it can and cannot prove.
Run the demonstrationThe proposition
A model demonstration asks whether one behaviour can work under chosen conditions. The Decision Foundry asks whether the complete product should advance. Customer value, intelligence, hardware, software, safety, operations and economics are evaluated together. A convincing output is not a pass: the evidence determines whether the decision is go, redirect, sequence or stop.
One connected system
The Product System governs why and when the product advances. The Physical AI system architecture governs how the system behaves in operation.
Product System
Strategy → Discovery → Productization → Launch → Scale
Five gates govern when the product may advance. Six decision lenses test whether the evidence meets the threshold for the current decision.
Physical AI runtime
Sense → Connect → Compute → Reason → Act → Orchestrate
Trust, safety and cybersecurity surround every layer.
Evidence return
Incidents, field variance, technical discovery or economics can reopen an earlier decision. Nothing advances automatically.
The visitor mission
Proposed lab zones
Follow one bounded initiative from declared intent to an inspectable decision record. As you move, the architecture resolves around what must be observed, authorized and proved.
Name the intended use, accountable owner, system boundary, acceptance threshold and stop condition before a model is selected.
Output: decision contractObserve the declared baseline, connect field signals to governed knowledge and compare options with sources, uncertainty and operating context visible.
Output: bounded proposalSchema, allow-list, policy and authority checks determine whether a proposal may reach deterministic control. Ambiguity or unsafe intent stops here.
Output: authorization recordRun one bounded action, inject declared variance, watch operator and system response, and prove that degradation, stop and rollback paths work.
Output: observed system behaviourCompare the result with acceptance and kill thresholds. Record the evidence, open assumptions and a go, redirect, sequence or stop recommendation.
Output: Product System gate recordIllustrative architecture
Change one condition. The model may interpret and propose; the system boundary determines what happens next.
Mistral-only runtime. EU endpoint. Owner-authorized.
The deployed environment permits deterministic code or exact, pinned Mistral models through a server-only EU endpoint. The production-allowlisted JARVIS, Search and Product Council paths are active by owner decision. Open assurance items remain disclosed; there is no global-region or provider fallback.
Live · deterministic
Economics, capacity, availability, constraints, risk propagation and hashes are computed by inspectable code with explicit assumptions, units and formulas.
Active · owner-authorized
The pinned moderation model must approve input and output. If moderation or EU inference is unavailable, AI fails closed while deterministic simulation stays available.
Active · owner-authorized
Small, Medium and Large serve as separate facilitator, systems-analyst and critic roles. Disagreement is preserved, and shared model-family failure modes are disclosed.
Local evidence
The VPS owns evidence identifiers, retrieval, prompt versions and state. No built-in web search, Files, Libraries, Agents or Conversations service is part of the production path.
Versioned evidence
Every substantive result records exact models, prompt and method versions, classifications, evidence IDs, deterministic hashes, limitations, review status and a human owner.
EU-only adapter
All hosted calls originate server-side at https://api.eu.mistral.ai. Other hosts, model aliases, unapproved IDs and provider fallbacks are rejected; secrets never enter browser code.
No model-to-machine or model-to-decision path
No Mistral response is a direct PLC, robot, charger or vehicle command. Deterministic code owns arithmetic and constraints; an independent safety function may reject or stop action; a named human owns investment, release, compliance, safety and customer commitments.
Explicitly outside production
OCR, embeddings, speech, Agents, Conversations, Files, Libraries, Batch, built-in web search, fine-tuning, Forge and preview services are disabled until the exact capability is verified for EU regional processing, Zero Data Retention, contract fit, security and measured need.
Hyperion Consulting is independent from Mistral AI. The EU regional endpoint is a technical boundary, not proof that the complete subprocessor and control-plane chain remains in the EU. No partnership, certification, sovereignty or compliance status is claimed.
Verify in official Mistral documentation
First mission packs
Each mission begins as planned and illustrative. It becomes measured only when the reference cell exists, the protocol is published and the result is reproducible.
Read machine state and a governed maintenance corpus, show supporting passages, state uncertainty and propose a bounded work order—without autonomously diagnosing equipment or issuing control commands.
Planned · illustrative until commissioned
Give the system an ambiguous or unsafe instruction. Intelligence may propose or abstain; typed interfaces, policy authority and independent safety determine whether anything can reach the controller.
Planned · illustrative until commissioned
Disconnect external connectivity and observe local inference, cached knowledge and deterministic fallback against declared degraded-mode requirements.
Planned · illustrative until commissioned
Challenge perception and learned behaviour with glare, occlusion, unfamiliar parts, drift and workstation variation to expose—not hide—the validated boundary.
Planned · illustrative until commissioned
Test a revision in simulation, release it to one controlled asset, expand only while acceptance thresholds hold and roll back when they do not.
Planned · illustrative until commissioned
Industry mission packs
The first three align with Hyperion’s published industry priorities. Smart infrastructure, logistics and defence are research contexts; publication does not imply sector client delivery.
Primary
Inspection, manipulation, cycle time, operator intervention, degraded modes and repeatable line deployment.
Primary
In-vehicle, edge and cloud boundaries; lost connectivity; staged OTA change; fleet observability and rollback.
Primary
Local autonomy, asset and fleet coordination, offline operation, incident recovery and cost to serve.
Research
Distributed sensing, multi-owner authority, privacy, cybersecurity, continuity and public-service constraints.
Research
Exceptions, human handover, mixed fleets, site variance, utilisation and repeatable deployment.
Research · bounded
Unclassified, non-kinetic and non-targeting missions using public or synthetic data. No weapons targeting or autonomous lethal decisions.
Evidence Passport
A demonstration is useful only when a buyer can tell what happened, under which conditions and what the result does not prove.
The record travels with the result
Question · intended use · data origin and rights · model, prompt, tool, firmware and dataset revisions · operating conditions · authority boundary · acceptance and kill thresholds · expected and observed result · failures · limitations · reproduction date · what the run does not prove
JARVIS · Lab OS
JARVIS lets a visitor observe system state, stress a bounded scenario, compare decisions and explain the result against cited evidence. It may synthesize and propose; it cannot waive an acceptance threshold, certify a system or authorize unsafe physical action.
Ask JARVIS about the FoundryAvailable now
These are current digital, bench and internal-R&D artefacts—not the planned physical facility. Each page states what it can and cannot prove.
One-image automation and safety observations; not a site audit, engineering assessment or certification.
Run the demonstrationA model describes what an industrial inspection layer might flag; not a validated defect detector or production-quality measurement.
Run the demonstrationSource-grounded general information about the EU AI Act; not legal advice or a conformity assessment.
Run the demonstrationA self-assessment across data, sensors, connectivity, ML readiness and governance; not an audit of your systems or a validated maturity benchmark.
Run the demonstrationA first-pass payback period and 3-year range from six inputs; not a quotation, a vendor recommendation or an engineering estimate.
Run the demonstrationA model reads a sensor CSV and drafts plausible anomalies and a schedule; not a trained maintenance model or a basis for maintenance decisions.
Run the demonstrationMeasured on the bench
AMD Ryzen AI Max+ 395 (Strix Halo) · 128 GB LPDDR5X-8000 unified memory · medians of three runs after a discarded warm-up · no tuning
qwen3:30b-a3b — 82.3 t/s · gpt-oss:20b — 38.6 t/s · devstral:latest — 14 t/s · mistral-small:latest — 13.8 t/s. These are configuration-bound observations, not a standardised benchmark or performance guarantee.
Read the complete field notesReachy Mini + SO-101
Measured · simulated · dry-run. Real sensing and teleoperation evidence stays separate from simulated autonomy and unexecuted robot actions.
Inspect the robotics evidenceAuralink
Pre-production · Hyperion-owned R&D. An inspectable reference implementation—not a client deployment or commercial-outcome claim.
How the complete system works
Select a stage to follow the system end to end.
01 / 08Device & Control
Sensors & environment. The physical world, sensed — cameras, depth, radar and OT signals.
See how we engineer this02 / 08Embedded Platform
Embedded platform. A real-time, deterministic, securely-updatable edge runtime.
See how we engineer this03 / 08Perception & Robotics
Perception. Mapping, detection and world-state resolved on-device.
See how we engineer this04 / 08Intelligence & Knowledge
Intelligence. Open-weight models and VLM/VLA policies propose an action.
See how we engineer this05 / 08Device & Control
Deterministic control. A deterministic controller turns a proposal into a precise command.
See how we engineer this06 / 08Trust, Safety & Cybersecurity
Safety & cybersecurity. An independent boundary accepts or blocks every command.
See how we engineer this07 / 08Perception & Robotics
Physical action. The validated command drives the machine in the real world.
See how we engineer this08 / 08Orchestration & Operations
Operations & fleet. Telemetry returns to operations; the system improves across a fleet.
See how we engineer thisOwned R&D portfolio
Hyperion-owned ventures at different maturity levels keep product, architecture and evidence questions concrete. Maturity labels are explicit; none is presented as a client deployment.
Pre-production reference implementation
Exercises: IntegrationReliability
Source toolkit under active development
Exercises: ArchitectureIntegration
Development prototype
Exercises: DataArchitecture
Source prototype with software and firmware boundaries
Exercises: IntegrationArchitecture
Development prototype
Exercises: ArchitectureIntegration
Development prototype
Exercises: GovernanceData
Development prototype
Exercises: ArchitectureData
Public early release
Exercises: GovernanceReliability
Development prototype
Exercises: GovernanceData
Demonstration platform
Exercises: GovernanceArchitecture
Decision instruments
The Foundry uses the same canonical Product System, architecture and decision tools as an engagement.
Framework / tool · released v0.1.0
Inspect verified fictional golden reports for changed evidence, assumptions, Release Cases and specialist conflicts. The viewer displays the pinned result; it does not recompute impact in the browser.
The system architecture from sensing to operations, surrounded by trust and safety.
Map deployment constraints to a hardware platform class.
A first-pass economics model for a robotics cell.
Test the system across architecture, data, operations and governance.
Inspect curated robot apps, datasets, policies and agents from the open ecosystem.
Commercial path
The Decision Foundry is an environment used inside Hyperion’s three existing mandates—not a fourth service category.
Make one consequential product decision with evidence. Receive an executive decision memorandum, evidence and risk map, and a prioritised 90-day plan.
Lead one Physical AI product from its current state to the next named evidence gate, with agreed artefacts, acceptance evidence and handover.
Embed fractional CPO or Head of Product leadership across product direction, architecture decisions, operating cadence and capability transfer.
Decision Foundry questions
No. The physical facility is in concept development and is not commissioned. Current digital demonstrations, measured bench work and Hyperion-owned R&D are labelled separately.
No partnership or special access is claimed. Hyperion is independent from Mistral AI. The owner-authorized JARVIS, Search and Product Council paths use a Mistral-only runtime through the EU endpoint, with deterministic fallbacks.
No. Model outputs are proposals that must cross schema, allow-list, policy and authority checks before deterministic control. Independent safety remains outside model authority.
No. It does not replace legal, safety, cybersecurity, conformity-assessment or certification specialists. It creates product and architecture evidence for a bounded decision.
A bounded decision record: observed evidence, unresolved assumptions, limitations and a go, redirect, sequence or stop recommendation.
Manufacturing, automotive and energy are primary contexts. Smart infrastructure, logistics and defence are explicitly labelled research contexts; publication does not imply sector client delivery.
Bring one stuck initiative
In a 30-minute fit call, Mohammed will identify the decision, the missing evidence and whether a Foundry mission or one of Hyperion’s three mandates is appropriate—or say plainly that it is not.
30 minutes · no obligation · Mohammed leads every conversation