A modern city operates tens of thousands of cameras and a control room staffed by a few dozen people. A port, a power grid, an airport perimeter — each generates more sensor data in an hour than its operators can review in a week. Humans were never going to watch all of it. The question was never whether machines would take over the watching. The question is whether they can be trusted with it.
FieldX was built for the battlefield. Our systems watch some of the harshest environments on earth — sub-zero altitudes, desert heat, coastlines where a vessel must be found without a transponder, airspace where the threat is small, fast, and uncooperative. They were designed for conditions in which sensors are jammed, data is spoofed, and an adversary is actively trying to deceive them. That work continues. This note is about why the same problems — and the same architecture — now matter far beyond it.
We think trust breaks into two deeply intertwined problems: intelligence and responsibility. Not one, then the other. Both, simultaneously, from the very beginning.
The intelligence required has three essential qualities.
A vehicle passes a toll gantry at highway speed — the system has a fraction of a second to read, classify, and charge. A drone crossing an airport boundary closes the distance to a runway in under a minute. A fault in a substation compounds while a video frame buffers to the cloud. This is not a software optimisation problem you solve with faster processors. It is an architectural constraint that shapes every decision: what hardware you choose, how your models are structured, and where computation happens — at the edge, on the pole, at the gate, not in a data centre three hundred kilometres away. Speed is not a feature. It is the medium in which infrastructure intelligence must exist.
A single sensor sees a sliver of reality. A camera sees light. A thermal sensor sees heat. An RF receiver hears signals. None of them, alone, understands what is happening. A stalled truck is just an object — until the system also knows the traffic pattern, the time of day, the weather, and what that stretch of highway looks like on an ordinary Tuesday. Real intelligence requires fusing every available source — electro-optical, thermal, radar, RF, acoustic, historical patterns, situational context — into one coherent picture. Not sequentially. Simultaneously. The system that can draw on every sense it has, in the moment it needs to, is the system that will make sound decisions.
Cities are not laboratories. Fog rolls in. Monsoon rain streaks every lens. Dust settles on optics; cameras age, drift, and fail; glare washes out the frame at exactly the wrong hour. And critical infrastructure has adversaries of its own — intruders who know where the cameras point, drones flown by people who would rather not be seen. The hard problem is not analysis when everything works. It is analysis when sensors disagree, when inputs are incomplete, when the scene is ambiguous. The system must make sound decisions under these conditions — not despite them, but through them.
If you can do it fast enough and comprehensively enough, with sound judgment when conditions degrade — you will appear intelligent. That is the intelligence we are building.
A system that is fast but reckless is a liability. It will escalate conflicts, cause collateral damage, and destroy the trust that makes autonomous systems viable.
Autonomous does not mean unaccountable. And this matters even more in a city than on a battlefield, because the people in frame are not combatants. They are commuters, workers, families. Responsibility here is not a constraint imposed on the system from outside. It is a capability the system must possess from within.
Not every anomaly is an emergency. A stray dog on a perimeter is not an intruder. A hobbyist’s drone over a park is not one loitering above a refinery. A stalled car is not a pileup. The system must classify with nuance and calibrate its response proportionally —because overreaction is as corrosive as underreaction.
Every false alarm teaches the operator to ignore the next alert. Alert fatigue is how surveillance systems die: not switched off in protest, but tuned out in exhaustion.
A system watching public space carries obligations a battlefield system does not. Privacy by design.
Data minimisation — process at the edge, transmit the event, not the footage.
Auditability — every automated decision traceable after the fact. And an understanding of second-order effects: a false alert that closes a highway at rush hour, an evacuation call that creates the very crowd crush it was meant to prevent. An algorithm that optimises for the immediate detection without considering what its response sets in motion is not intelligent — it is merely reactive.
This may be the hardest capability of all. The system must possess enough self-awareness to recognise when it is out of its depth — when the scene is ambiguous, when its confidence is low, when the stakes demand human judgment. It must escalate appropriately, presenting the decision to an operator with full context and its own assessment. This is not human-in-the-loop as a compliance checkbox. It is human-in-the-loop as an AI problem: the system needs enough intelligence to know what it doesn’t know. Building a system that can detect an intruder is hard. Building a system that knows when it is uncertain whether something is an intruder — and acts differently because of that uncertainty — is a fundamentally deeper problem.
Operators set the policy, the thresholds, the rules of response. Machines execute within those bounds. When the bounds are unclear, the machine asks. This is not a mode you can toggle on or off. It is architecturally embedded — designed into the system from its first line of code. The machine’s sense of autonomy is never without the human. It is autonomy in service of human command, not in place of it.
You cannot build intelligence first and bolt on responsibility later. A system that is fast but indiscriminate drowns its operators in false alarms and is quietly abandoned within a season — personnel revert to watching the screens themselves, and the investment becomes furniture. A system that is careful but slow is simply irrelevant: the vehicle has passed, the drone has landed, the incident has already happened.
The architecture must be designed from the start to hold both in tension. Speed and judgment. Comprehensiveness and restraint. Autonomy and accountability.
Most companies entering this market arrive from the commercial side — a camera company adding analytics, a software company renting someone else’s sensors. FieldX arrived from the opposite direction. We built for the most demanding customer on earth first. Our systems were designed for soldiers at −40°C and for deserts at +60°C, for coastlines and contested airspace. FX-Edge taught us real-time inference under the most extreme size, weight, and power constraints imaginable. FX-Ground taught us persistent, autonomous awareness across mountains, jungles, deserts, and coasts — terrains that each break your models in different ways. FX-Sky taught us to track fast-moving, coordinated aerial objects at city scale, fusing RF, radar, and electro-optics. Defence is an unforgiving teacher. It never permits the assumption that conditions will be good, that sensors will be clean, that the subject will cooperate. A tolling gantry in monsoon traffic, a port perimeter at three in the morning, a drone near a stadium — these are hard problems. They are also, structurally, the problems we have already been solving, in environments deliberately designed to make them harder. In one sense, FieldX is not entering a new market. We are removing a uniform. And the deepest lesson transfers whole. The individual smart camera is becoming a commodity. The individual analytics dashboard is becoming a commodity. What is not a commodity — what almost no one is building — is the connective tissue between them: one intelligence layer that fuses the street, the perimeter, and the airspace into a single operating picture, and knows when to hand that picture to a human.
We are not building isolated products. We are building one connective tissue, accessed through many interfaces.
This way of seeing the problem — as a system of systems, a network of connected nodes rather than isolated devices — comes directly from the founder’s background. Kaustuv’s work spans AI and network science, the discipline of understanding complex systems as interconnected graphs of relationships. His previous company, Vibrant Data, was built on the premise that complex scenarios become legible when you see them as networks. A city is, quite literally, a network — of roads, of flows, of infrastructure, of the relationships between them. The mental model FieldX brings — seeing the whole, not just the parts — was made for this.
