Sensor & platform networks
Build and operate the physical supply — satellite constellations, drone fleets, monitoring and camera networks.
SupplyDiscovered and booked as resources that provide a capability.
THE CAPABILITY OPERATING SYSTEM
The physical world,
programmable.
Tilfang is designed to turn one sentence of intent into a governed mission across drones, sensors, satellites and analysts — so the how is no longer your problem.
Concept & architecture — to be validated through a simulation-first pilot
The concept, in numbers
Design counts, not traction — nothing here claims usage.
The fragmentation problem
The capabilities already exist. Access to them does not.
Sensors, platforms, APIs, operators, and analytical services are distributed across incompatible systems and organisational boundaries. Completing one mission often requires specialist knowledge, manual coordination, and multiple interfaces.
What the world already has
- Cameras & telescopes
- Satellites & remote sensing
- Drones & robots
- Ships & aircraft
- Weather & environmental sensors
- Laboratories & instruments
- Industrial & inspection systems
- Specialist human operators
- Data services & analytical APIs
21.1 billion connected IoT devices estimated worldwide by the end of 2025, growing 14% a year — and connected devices are only one slice of the capability the world already operates.IoT Analytics · State of IoT 2025
What one mission demands of you
- Knowing which systems exist
- Knowing who owns or operates them
- Knowing how to access them
- Knowing which interfaces they expose
- Knowing whether they are currently available
- Knowing which regulations and permissions apply
- Knowing how to combine several systems
- Knowing how to evaluate mission success
The inversion
Start with the mission, not the machine.
Current automation starts with a known system and asks what can be done with it. Tilfang reverses that model: start with the objective, derive the capabilities, then identify and orchestrate the resources.
Traditional approach
- Choose device
- Operate interface
- Collect output
The device defines what is possible. The human does the integration work.
Tilfang approach
- Define objective
- Derive capabilities
- Find resources
- Govern & execute
The objective defines what is needed. The system finds, compares, and coordinates the resources — under policy.
Tilfang builds no sensors, operates no fleet, and trains no perception models. It is the orchestration layer above the capability networks others build — not a competing supply network.
What this layer is not
Perception & analysis models
Turn raw signals into meaning — imagery, time series, and multimodal streams read by a model rather than a person.
SupplyBooked against a capability node like any other resource.
Device control & robotics middleware
Command individual machines reliably and safely at the device level.
SupplyReached only through a certified adapter, never driven directly.
Enterprise data & agent platforms
Integrate business data, or plan and execute workflows over software tools.
AdjacentThe same planning instinct, applied to records and APIs rather than physical capability.
The layer would begin where these stop: deciding which capability an objective actually requires, which resource should provide it, and under whose policy and approval it may run.
Interactive demo
See a mission unfold.
Three sample objectives, decomposed step by step: the capabilities they require, candidate resources compared on reach, lead time and approval, the constraints that decide the selection — and the evidence record the mission would leave behind.
- Inspect infrastructure
- Diagnose an industrial anomaly
- Assess an environmental area
Open the demo
Conceptual demo · synthetic data · no live devices
Platform
One capability layer between intent and evidence.
The full design lives on its own page: the abstraction, the layers that govern it, the path into it, and the autonomy it refuses to assume.
- Capability graphcapability nodes and example resource classes, each declaring reach, lead time and approval
- Seven architecture layersplanning never equals permission; permission never equals execution
- Capability onboardingthe ten-step path by which an unknown resource would become bookable — and stay revocable
- Trust & autonomyAI proposes, policy constrains, humans remain accountable — autonomy is granted in levels, never assumed
The plan
A narrow pilot first. The network later.
How the universal vision becomes a first product, what the money would attach to, and the phases in between — each on its own page.
- Initial pilota simulation-first, civilian drone observation and inspection pilot — its scope, its mission, and what stays out
- Business modelenterprise deployments first; a unit, a precondition and a stated weakness per stage
- Roadmapfrom one bounded mission to a governed multi-provider capability network
The experiment
Deliberately falsifiable.
A concept that cannot fail is a brochure. This one states its tests — and what each outcome would mean.
The schema bet
Published 2026-08-13, before any product: the capability vocabulary, public and versioned. An operator either describes a real asset in it — or argues with it. Both outcomes are findings.
The pilot bar
One bounded, simulation-first mission with a design partner has to show an objective becoming an explained, governed, auditable plan — before anything real is ever flown.
The kill criteria
If operators cannot describe their assets in a shared vocabulary, or no compliance owner will sign one narrow scope, the thesis fails — and this site will say so.
Why now
AI is the catalyst. Architecture and trust are the foundation.
Several developments make capability orchestration newly plausible — but AI progress alone does not solve the product. The hard problems are architectural, legal, and commercial.
What has changed
- Foundation models interpret natural-language objectives
- Agentic systems plan and select tools
- Multimodal models analyse sensor outputs
- Structure can be extracted from API documentation
- Digital twins and simulators allow safer testing
- Robotics & IoT ecosystems expose ever more interfaces
- Policy engines and identity systems govern machine actions
- Marketplaces exist for data, compute, imagery, and services
What stays hard
- Trusted capability descriptions
- Reliable, verified adapters
- Permissions and ownership
- Safety and legal constraints
- Data quality and availability
- Commercial agreements
- Verification of real-world outcomes
Sources · 8
Each enabler is a general statement about the state of the technology, not a claim about this concept. The source given is one current primary example that the development is real — not an exhaustive survey.
- Foundation models interpret natural-language objectivesOpenAI · Function calling guide
- Agentic systems plan and select toolsarXiv · TRAJECT-Bench, an agentic tool-use benchmark
- Multimodal models analyse sensor outputsNASA · Prithvi, a geospatial foundation model in orbit
- Structure can be extracted from API documentationSpeakeasy · Generating MCP tools from OpenAPI
- Digital twins and simulators allow safer testingNVIDIA · Isaac Sim robotics simulation
- Robotics & IoT ecosystems expose ever more interfacesIoT Analytics · State of IoT 2025
- Policy engines and identity systems govern machine actionsCNCF · Open Policy Agent
- Marketplaces exist for data, compute, imagery, and servicesSnowflake · Marketplace
Defensibility
The moat is not the language model.
Concept stage · nothing accumulated yet
Models improve and become commoditised. Defensibility would come from what accumulates around them: verified knowledge, trusted integrations, and operational history. None of it can be bought in one move — most of it is earned mission by mission, adapter by adapter, provider by provider, and the rest depends on people outside the company entirely.
Capability graph
A high-quality semantic model of which resources provide which capabilities under which conditions.
How it would be earnedGrows whenever a real mission forces a capability to be described precisely enough to be matched, priced, and governed.
Verified adapter network
Integrations into real systems and providers — tested, versioned, and cleared for a defined scope.
How it would be earnedOne adapter at a time: sandbox, simulation, human or provider verification, then certification limited to that scope.
Policy & trust layer
Reusable governance patterns for lawful, safe cross-organisational execution.
How it would be earnedAccumulates from each approval negotiated with a real legal, safety, or compliance owner — and reused on the next mission.
Execution history
Knowledge of how resources actually perform across conditions, missions, and locations.
How it would be earnedOnly from missions that ran. This one cannot be licensed, scraped, or generated — it is the slowest asset on the list.
Provider network
Relationships with sensor owners, operators, data providers, and service companies.
How it would be earnedBuilt provider by provider, each with commercial terms and a capability description someone is willing to stand behind.
Mission templates
Proven decomposition and planning patterns for specific mission types.
How it would be earnedDistilled from missions that ran end to end — including the ones that failed, and the reason they did.
Two that would compound elsewhere
The six above accumulate inside the platform, mission by mission. Two further forms of defensibility would form outside it, on mechanisms worth naming separately rather than filing under the same heading.
Workflow embed
Not a dataset but a position: the capability layer wired into one organisation’s identity system, its approval chains, and the audit record its auditors already read.
What it depends onDeepens per customer rather than per mission, and only with customers willing to let a new layer near those three things. It is also the shallowest of the eight — an incumbent can build the same integration — so depth here buys a lead, not a lock.
Standard-setting potential
If the way a capability and a mission contract are described here becomes a vocabulary other people reuse, the schema carries weight even where the platform is absent.
What it depends onAdoption by parties with no stake in this concept — the only item on this page that cannot be earned alone. It is also the one testable before any platform exists: publish the capability schema, and see whether an operator can describe a real asset in it, or argues with it.
None of these assets exist today. What can be judged at concept stage is whether they are the right ones to accumulate, and whether the initial wedge is the cheapest honest way to start.
“Certified” on this site means one thing only: the concept’s own onboarding path has cleared a specific adapter for a named capability scope, and can revoke that clearance. It is not a regulatory or third-party safety certification, and none is claimed. Capability certification sits in roadmap Phase 3, after the controlled pilot.
Where this sits
Adjacent markets, not a market size.
A top-down market number at concept stage is a guess presented as a forecast. What can be stated honestly is where this work is already being done and paid for, and which seam a capability layer would attach to. These are the six markets the concept sits adjacent to.
Industrial inspection
TodayRecurring — and often legally mandated — condition assessments of built assets: structures, plants, energy and transport infrastructure. Each round is scoped, scheduled, and evidenced largely by hand.
The seamA repeating objective compiled into a governed plan, with the evidence trail produced by execution rather than assembled after it.
Remote sensing & earth observation
TodayImagery and derived data sold as archive scenes, new tasking, or subscriptions. The buyer picks sensor, resolution, and product before the question is answered.
The seamA request stated as a capability and a quality bar, with the product selected to satisfy it instead of chosen up front.
Drone operations & services
TodayLicensed operators fly bounded missions under an authorisation granted by the national aviation authority, tied to that specific operation and the risk it carries.
The seamOperators as bookable capability providers, with authorisation and approval carried as mission constraints — never routed around them.
IoT & robotics orchestration
TodayFleets, devices, and machines are managed per vendor and per platform, each with its own control plane, data model, and permissions.
The seamOne capability abstraction above those control planes, so a mission can combine several without an integration project per pair.
Environmental monitoring
TodayAgencies, utilities, and consultancies run repeat measurement programmes where comparability across seasons and years matters more than any single reading.
The seamMission templates that record which resource was used under which conditions, so a later run stays comparable to an earlier one.
Mission & field-operations software
TodayPlanning, dispatch, and reporting tools organise the people and jobs of a single organisation, around resources it already owns.
The seamThe same workflow composed across organisations and providers, and grounded in what a resource can actually do.
The initial wedge sits where the first three overlap: a documented visual and thermal assessment of a defined, authorised area is one bounded mission whose supply chain already exists — an operator, a sensor, an analysis step, a report — which is what makes it the cheapest honest test of the abstraction.
Adjacency says where the concept would compete for a budget that already exists. It is not evidence that anyone wants it — that is what the pilot is for.
Sources · 3
Three of the six describe regulation or commercial practice that can be checked; the rest describe how tooling is organised today. Sources are given for the first three. None of them says anything about demand for this concept.
- Periodic condition assessment of built assets is in places legally mandated, not merely good practice.EU · Road Infrastructure Safety Management Directive (2019/1936)
- Earth-observation imagery is sold as archive scenes, new tasking, or subscriptions.Geoawesome · Demystifying satellite data pricing
- A commercial drone mission in the ‘specific’ category runs under an authorisation from the national aviation authority, tied to that operation and its assessed risk.EASA · Operational authorisation, ‘specific’ category
FAQ
Direct answers to fair questions.
The questions a sceptical reader actually asks — answered directly, on their own page.
- Is Tilfang an autonomous AI agent?
- Does Tilfang replace device-control systems?
- Can Tilfang use resources it has never seen before?
- Why focus on capabilities instead of devices?
- Why start with drones?
- How is Tilfang different from sensor networks, analysis models, operational data platforms, robotics middleware, or AI-agent frameworks?
- What is the long-term opportunity?
Founders
Space systems meet cybersecurity.
Two ETH Zürich backgrounds, one shared conviction: the hard part of orchestrating the physical world is not the AI — it is trust, governance and systems that must not fail.
Nebe
Space systems
ETH Zürich
Dina
Cybersecurity
ETH Zürich
Full attribution follows separately — first names until it lands.
“The next generation of AI should not stop at answering questions. It should help people coordinate the capabilities already present in the world — safely, transparently, and across system boundaries.”
None of this is settled at a desk. Whether an operator can describe a real asset in a shared vocabulary; whether a compliance owner will sign one narrow scope rather than refuse a broad one; whether a plan survives the constraints of an actual site — those answers exist only inside somebody’s operation. That is why the next step is not more architecture. It is one real mission, with the people who already run it.
Design partners
Build the first mission with us.
The first pilot will be shaped with a small number of design partners, each around one bounded, real mission they already run.