In a building of thirty years or more, the as-built plumbing drawings are often missing, incomplete, or overtaken by decades of alteration. The system does not depend on them. A Method-of-Characteristics 1D transient flow model builds a hydraulic digital twin of your network from coarse building parameters — storey count, flats per floor, riser material and diameter, meter room positions and the hydraulic zoning arrangement. From that twin we generate synthetic transient scenarios, over 220,000 of them, covering leaks of different sizes at different positions, and the model is trained on those. Topology resolution can be set anywhere from coarse to fully detailed: better information about your building produces a better twin, but the absence of drawings is not a blocker. Once the sensors are on site, the model is fine-tuned to your actual network by transfer learning.

Every time a tap opens or a valve closes, a pressure wave travels through the network and reflects off every junction, fitting and change of section it meets. A leak changes those reflections in a way that is consistent and physically predictable, and that is what the model is trained to recognise. The training set is the 220,000 synthetic scenarios, which include the network with and without leaks under a range of transient events; a hierarchical multi-task 1D-CNN learns the difference and is then fine-tuned on your building, so it is calibrated against your network's own normal behaviour rather than a generic one. Under simulation the model retained 96.05% detection accuracy when the signals were heavily corrupted with noise, at a signal-to-noise ratio of 5. What we cannot yet give you is a false-alarm rate measured over months in a fully occupied tower. Establishing that under real occupancy is the specific purpose of the live pilot phase in our roadmap, and we would rather say so than quote a number we have not measured.

The methods in use today fall into two groups, and each has the same structural weakness. Fluorescent dye tests, colour water tracing, moisture meters and acoustic listening are accurate, but they need a technician working inside the affected flat — and residents routinely refuse access, which is a large part of why more than 35% of Hong Kong seepage cases close without the source ever being found. Microwave surveys and infrared thermography are non-intrusive, but they detect damp in a structure: they tell you moisture is present somewhere behind a surface, not which pipe is losing water. Both groups also produce a snapshot taken on the day of the visit. EcoFeatures measures the pipe network itself, from the communal duct and meter room, continuously — so an intermittent leak that would be dry on the morning of an inspection is still caught, and the answer is a location in the pipe network rather than an area of damp on a wall.

No. Every sensor sits in the pipe duct or the meter room — communal areas your staff already hold keys to. There is no access negotiation with residents, no appointment to arrange and nothing to refuse. This matters beyond convenience: one of the four outputs of each inference pass is whether the leak sits inside a unit or in the communal pipework outside it, so if entry does eventually become necessary, you are asking one household for access to one flat with a reason, rather than canvassing a whole stack.

Dual-mode edge nodes, fitted in the pipe duct and the meter room. Mode one is an external clamp-on triaxial accelerometer, which straps to the outside of the pipe and touches nothing wetted. Mode two is an in-line pressure sensor, used where a suitable tapping point exists and the building permits it; a wetted fitting of that kind must carry Water Supplies Department General Acceptance certification. Which mode suits your building is decided at the site survey. Both capture at 1 kHz. Nothing is emitted into the pipe: the system uses only the transients your residents already generate by opening taps and closing valves, so no test wave is injected and no water is discharged for testing. Meter rooms are the natural measurement point because they are the distribution nodes of the network — a meter room can house up to six meters and serve a floor, and sometimes the floor below it as well.

A site survey comes first, producing a sensor plan for the riser and the meter rooms. Installation itself is plug-and-go: clamp-on nodes fitted in the communal duct and meter rooms, with no pipe cutting, no wall opening, no drilling and no interruption to the water supply. It takes hours rather than days, and residents need not be asked to stay in or stay out. The rest of the deployment runs after that — the synthetic hydraulic model of your network, on-site fine-tuning of the AI by transfer learning, and a system baseline with a network condition report. The complete package is delivered within a month of the survey.

They come from two different exercises, and they should never be read as the same thing.

Simulation. Across the 220,000 synthetic scenarios, a single-task 1D-CNN predicted the correct floor 99.59% of the time and the correct individual flat 99.44% of the time, under clean signal conditions. Under heavy noise corruption, at a signal-to-noise ratio of 5, detection accuracy held at 96.05%. These are model results on modelled data. They set an upper bound, not an expectation.

Physical pilot. At HKUST Tower 4, measured on a real working network: 98.8% leak detection accuracy using combined pressure and flow sensing on valve-closing events, 97.5% from a pressure sensor alone and 96.3% from pressure plus vibration. The full measured range across all sensor configurations and transient types was 86% to 99%. The 86% end of that range is as real as the 99% end, and which end you land nearer depends on which sensors are fitted and which kind of transient the reading is taken from.

We keep the two sets apart in every document we publish. If you ever see one of our figures without the word simulation or pilot attached to it, ask us which it is.

The system was modelled and tested over a leak range of 0.01 L/s to 0.1 L/s, and 0.01 L/s — roughly a persistent weep at a joint — is the smallest rate tested. So the honest phrasing is "as small as 0.01 L/s", not "below 0.01 L/s"; we have no evidence for anything finer and do not claim it. What the system will not do is give you a millimetre position along the pipe. The output is hierarchical: leak present, floor, flat, inside or outside the unit. For scale, a single Type A apartment at the pilot site contained more than 134 individual pipe segments and fittings, with runs up to 47 m — so a floor-and-flat answer narrows a search that would otherwise cover an entire stack, but a plumber still opens the last metre.

Not a number, until we have taken a baseline on your network — and you should be sceptical of any supplier who gives you one before that. Every measured figure we publish comes from one building: HKUST Tower 4, twelve storeys, three flats per floor, a copper riser, three hydraulic zones. Your building is unlikely to match it. The mechanism that carries the model from Tower 4 to a new network is on-site transfer learning, which re-tunes the model against your building's own measured transients, and the digital twin is rebuilt from your parameters rather than reused. How well that transfers to a building materially different from the pilot is precisely the question a pilot answers. What we can commit to beforehand is the method, the deliverables and the baseline report — not a percentage.

Partly, and here is the boundary. The pilot building is twelve storeys with three distinct hydraulic zones — floors 1 to 4 supplied directly from the riser, floors 5 to 8 gravity-fed from the roof tank, and floors 9 to 12 served by a circulation pump through an intermediary middle tank. Multi-zone behaviour, break tanks and pumped circulation are therefore inside the model's scope in principle, and the Method-of-Characteristics model extends to taller networks without a change of method. But Hong Kong has roughly 8,000 high-rises and more than 1,500 buildings over 100 m, and a 40- or 60-storey tower has more zones, longer risers and more transfer plant than anything we have physically instrumented. For those buildings we have a modelling case and not yet a measurement, and we will say that to your committee as readily as we say it here.

Material matters to the physics, because pipe material, diameter and wall thickness set the speed at which a pressure wave travels and how it reflects. It is a parameter of the digital twin, captured at the site survey — the model is built for the pipe you have, not for a generic one. What we have not done is validate physically on anything but copper. The Tower 4 riser is high-grade copper, 50 mm internal diameter with a 2 mm wall; floor distribution branches are 28 mm copper and in-flat pipework 26 to 28 mm. Galvanised iron, stainless steel and plastic risers are all modellable and all common in the 30-year-old Hong Kong stock we are built for, but on those materials we would be taking a first physical measurement with you rather than repeating a proven one.

Not directly, and this is worth settling before a pilot rather than after. The method reads pressure transients travelling through the pressurised fresh-water supply network, so that network is what it diagnoses. A drainage or waste stack is not pressurised and carries no such transients; seepage from a failed waterproofing membrane, defective tiling or a floor slab is not a supply-pipe leak at all, and the system will not find it because there is nothing in the supply pipe to find. That said, part of what makes seepage investigation slow and expensive is that nobody knows which of those causes they are dealing with. A building whose supply network is under continuous monitoring can rule the supply pipes in or out before the dye tests and the access requests begin — a real narrowing of the problem, but not the same as finding every seepage source.

No, it is not a finished product, and you should price that in. We are at TRL 4 to 5 — validated in a living lab, which is the HKUST Tower 4 pilot. Our target is TRL 7, a system prototype in an operational building, by the middle of 2027, through three phases: digital twin development from September to December 2026; AI module development with transfer learning from the Tower 4 pilot, September 2026 to March 2027; and live pilot implementation in an occupied operational building, January to August 2027. A building that engages with us now is a development partner rather than a customer buying a mature system, and the commercial terms, the expectations and the reporting should reflect that. The core method is covered by a United States provisional patent filed on 1 December 2025 — IP.PA.12794, application 63/938125 — which is a provisional filing, not a granted patent.

From us: a site survey and sensor plan for the riser and meter rooms; plug-and-go installation of the edge nodes; a synthetic hydraulic model of your network; on-site fine-tuning of the AI by transfer learning; and a system baseline with a network condition report. That deployment is completed within a month. Continuous monitoring then runs as a subscription with a minimum term of twelve months, which is what gives the model enough of your building's ordinary behaviour to be useful.

From you: access to the pipe ducts and meter rooms, the coarse building parameters we need to build the twin, a point of contact in your maintenance team, and agreement on how findings are handled when the system flags a flat. The Tower 4 pilot was run in coordination with the HKUST Campus Management Office and in compliance with Water Supplies Department regulations; we would expect the same coordination with your building services team and, where relevant, your Owners' Corporation.

We will also agree beforehand what counts as success. That is easier to settle honestly at the start than to argue about at the end.

One completed pilot, and no signed customers. We would rather say that plainly.

HKUST Tower 4 is a completed living-lab pilot, and it is the source of every measured figure on this site. Peak Tramway Limited has asked us for a formal proposal, to be tabled to their senior management for pilot approval at St. John's Building and The Peak Tower — that is a request for a proposal, not a contract, and nothing has been installed. Sino Group's building management team has confirmed the problem for their 30-plus-year portfolio, where the loss falls on the fresh-water supply rather than the cheaper seawater flushing supply — again, a confirmed pain point, not a customer.

We have no commercial deployments, no case studies beyond Tower 4 and no deployment count to quote. The institutional support behind us is the HKUST Office of Knowledge Transfer, the HKUST Entrepreneurship Center, the HKUST Campus Management Office and the Hong Kong Innovation and Technology Commission, and we are an HKUST Techship grantee under the Lo Kwee Seong Fund.

The Tower 4 pilot was conducted in compliance with Water Supplies Department regulations and coordinated with the HKUST Campus Management Office. Two points matter for your building. First, where the deployment uses external clamp-on accelerometers, nothing wetted is touched at all — the sensor straps to the outside of the pipe, so your plumbing installation is unaltered. Second, where in-line pressure sensors are used, the fitting is a wetted component and must carry Water Supplies Department General Acceptance certification; pipework in these systems meets the relevant BSEN standards for potable water. No pipe is cut, no existing fitting is replaced and the supply is not shut down for installation.

What is recorded is hydraulic signal: pressure and vibration time series sampled at 1 kHz from sensors in the communal riser duct and meter rooms. There is no camera, no microphone in any living space and no device inside anybody's home — every node sits outside the residential unit, on communal pipework. The system is trained to classify leak signatures; it is not a consumption meter, it does not replace or duplicate Water Supplies Department metering, and it produces no household billing record. The one privacy-relevant output is that an alert names a floor and a flat, because that is the location of the suspected defect: a location in the building, not a record of what anyone did.

On ownership and retention we have no published standard policy to point you to at this stage, and we will not pretend otherwise. Those terms are settled in the service agreement for each deployment. If you are considering a pilot, expect data ownership, retention period, access rights and what happens to the data if either party withdraws to be written down before any sensor is fitted — ask us for that in writing, and do not accept a verbal answer from us or from anyone else.

A dashboard, and alerts. When the system flags a defect you receive a real-time micro-leak alert carrying the location as Floor, Flat and inside-or-outside the unit, together with a severity grading, so a slow weep and a developing failure are not presented to you as the same event. Alongside that sit predictive asset-lifecycle analytics on the network's condition and quarterly performance reporting you can put in front of a management committee or an Owners' Corporation meeting. The one-time survey also produces a system baseline and network condition report, which is often the document that is useful first — it describes the state of the pipework you are budgeting for. For portfolios already running a building management system, API integration is available as a separate item.