Commercial vehicles rarely fail without warning. The clues are often buried in thousands of data points, each seemingly unremarkable on its own. For fleet operators, finding those early signals can mean the difference between a planned repair and an expensive roadside breakdown. Intangles is using artificial intelligence grounded in vehicle physics to make sense of that complexity across more than 500,000 vehicles. In this conversation, Neil Unadkat, CTO and co-founder of Intangles, explains how the platform identifies emerging faults, helps fleets reduce downtime and fuel costs, and points towards a future in which vehicles can take on more maintenance decisions themselves, autonomously.

We monitor over 450 signals across a vehicle’s engine, aftertreatment, transmission and electrical systems. Across the 500,000+ vehicles on our platform, that is more than 15 trillion data points a year, and on any given day, nearly all of it says the vehicle is fine. Most fleets still rely on fault codes, and a fault code usually appears only after the damage has already started. The earlier warning sits in readings that each look normal on their own.

Finding it starts with deciding what normal means for a specific vehicle at a specific moment. A coolant temperature of 100°C on a fully loaded truck climbing a ghat in May is fine, while the same reading on that truck idling at a depot is an early sign of trouble. Every reading is judged against the operating envelope we expect for that engine class at that engine speed, load and ambient temperature.

Even a reading outside the envelope is usually a loose connector or a sensor glitch. Wear in a real component shows up in several related signals at once. When a turbocharger starts to deteriorate, boost pressure, intake airflow, exhaust temperature, and fuel consumption shift together as the engine’s physics predicts, and we act on that combined pattern. Comparing the vehicle with others of the same class doing similar work then tells us whether the drift belongs to that one truck or to the model, and which causes are most likely.

A fleet manager running forty trucks needs to know which two to pull in this week. I think any alerting system should be judged on whether it gives them that answer.

Why does combining vehicle physics with AI make predictions more useful than relying on AI alone?

Mostly because commercial vehicles produce very little usable failure data. A component can run for years before it fails, service records usually say what was replaced and rarely why, and the same fault looks one way on a mining tipper and another on a city bus. A model that learns only from past failures has little to learn from, and it learns slowly.

How oil pressure should respond to engine speed, or how an alternator’s output should change with RPM, is well-understood engineering. When we build those relationships into the model, it knows what healthy behaviour looks like from the first day it sees a vehicle. Machine learning then adjusts to that individual vehicle as more data comes in. This matters most where the industry has the least history. Electric vehicles and the aftertreatment systems that arrived with BS-VI have been on Indian roads for only a few years, too short for a purely data-driven model to have seen many of their failures.

The physics also lets us explain a prediction. Voltage fluctuations can come from several different electrical faults, and each behaves differently as engine speed changes. By modelling the alternator’s turns ratio, we can tell them apart, so the workshop knows which part to replace. A purely statistical model would flag the anomaly and leave the technician to work out where it came from.

This is how we hold predictive accuracy at 96%. Each wrong prediction pulls a revenue-earning vehicle off the road for nothing, and operators notice that cost within weeks.

What changes for a fleet operator when a vehicle can warn you about a potential problem well before it becomes a breakdown?

In my experience, most commercial fleets in India still maintain vehicles on a fixed-kilometre schedule or only after something breaks. Fixed schedules replace parts that still have life left and miss early failures. A breakdown costs far more than the repair. A truck that stops on a highway means towing, a late consignment, possibly a contractual penalty, a driver stuck on the roadside and often several days off the road waiting for parts at a workshop the operator has never dealt with.

With enough warning, the same repair gets booked for the end of a trip, at a workshop the fleet trusts, with the part in stock and the cost known in advance. It also changes how vehicles are assigned. An operator who knows a vehicle has a developing problem can keep it on shorter routes until it is serviced and send a healthier one on the long-haul run. Across our customers, this has meant a 75% reduction in powertrain breakdowns and a 10 to 30% increase in fleet uptime.

Emissions systems show why the timing matters. If DEF dosing drifts and the SCR system stops controlling NOx properly, the engine can derate and lose torque mid-trip. If caught early, the same problem can be fixed at the depot with an injector or sensor replacement.

Fuel efficiency gets less attention. A clogged air filter or fuel rail pressure drifting below its ideal range can increase diesel consumption well before either one stops a vehicle, and fixing these early is part of how fleets on our platform improve fuel efficiency by 2 to 10%. For most operators, fuel is one of the highest running costs, so even the lower end of that range is a real saving.

Looking ahead, how far can this technology go? Could commercial vehicles eventually move from simply reporting their health to actively making decisions about maintenance, routes, fuel, or charging?

I think a good part of it will, and some of it already has. The first decisions to move to the vehicle will be routine ones where a wrong call is cheap and a person adds little. Deciding when a diesel particulate filter should regenerate, and in which mode, based on how and where the vehicle is running, is one example. Another is choosing which charging station an electric bus should stop at, given its state of charge, realistic remaining range and the connectors available on the route. Fleet platforms can already recommend both, and I expect them to act on more decisions like these without a person signing off each time.

How far this goes depends on how good the underlying predictions are. The distance-to-empty figure on an electric vehicle’s dashboard often ignores payload and how hard the air-conditioning will have to work in the afternoon heat. A charging decision based on that number can leave a bus stranded, so we model range based on operating and environmental conditions before recommending where to charge.

Decisions with safety or regulatory consequences are different. Taking a vehicle out of service or overriding a driver should remain with a person for a long time, even when the system advises otherwise. Part of that is accountability, since the operator answers to a customer or a regulator for those calls, and part of it is trust, which a system earns through a track record. I expect operators to hand over more as that record builds, fleet by fleet and at very different speeds.