FORE — Software · foreKEY

Software · foreKEY

The tools we wished
existed.
So we built them.

Years of projects on water networks taught us that certain problems keep coming back, regardless of context. Data quality: measurement instruments fail, drift, or return implausible values for many reasons — sensor technology limitations, non-ideal installation conditions, mechanical stress, electrical interference, harsh environments. Software accessibility: hydraulic modelling is locked behind tools that are technically complex, hard to adopt, and out of reach for much of the engineering workforce that needs them most.

We built foreKEY to solve these problems.

The architecture

An operational tool suite, one philosophy.

foreKEY is a browser-based suite where each module pairs a deterministic engineering backend with an AI agent that orchestrates it. The backend holds the methodology we have built over years of field work. The agent acts as a smart colleague who knows the methodology and walks the engineer through it conversationally.

The AI reasons, the engine computes, the professional decides.

REED harmonises data from any source. NEWT brings hydraulic modelling to natural language. BLEAK cleans and reconstructs signals from sewer instruments. A third tool — water treatment — is in active development.

Distribution

NEWT

AI-agentic hydraulic modelling built on the US EPA EPANET 2.3 engine. Walks the pipes.

Sewer

BLEAK

Advanced anomaly detection, signal cleaning and reconstruction.

Data Harmonisation Layer

REED

Domain-agnostic. Reads the network. Temporal alignment · Small gap filling · Outlier detection · Unit conversion — the trusted data foundation for the entire suite. Scalable to large instrument fleets: the same pipeline runs on 10 sensors or 1,000, with no manual intervention per sensor.

No black-box AI

The agent orchestrates — selects which engineering tool to call, sets parameters, asks for confirmation when an operation is destructive. Every number comes from audited, deterministic engineering code. No hallucinated pressures, no invented flowrates.

Human-in-the-loop

Every destructive operation requires explicit confirmation. Full audit trail: every change tracked, every step reversible. Session recovery after a crash. The professional is always in control.

Browser-based, zero installation

Nothing to install, no plugins, no version dependencies. Works on desktop and tablet. The professional opens a URL and the tool is there. On-premise deployment available on request.

Your data stays yours

Raw time-series data never reaches the AI model. The agent works on aggregates and diagnostics derived from the data, not on the raw records. EU-hosted infrastructure. GDPR and NIS2 compliant.

Specialises over time

Confirmed user decisions feed a case library. The longer you use the tool, the better it performs on your network.

Frontier reasoning

Built on Claude by Anthropic. Concrete causal reasoning: the agent reads topology, weighs alternatives, tests hypotheses, proposes specific actions backed by arguments the professional can challenge.

REED

Reads the network.

Data Harmonisation Layer

Water networks generate heterogeneous measurements — different instruments, different sampling rates, different units, different column names. Before any analysis can begin, that data needs to be harmonised. REED is the data harmonisation layer that takes raw data from any source and makes it comparable, clean, and continuous.

REED applies to any kind of instrumental data — whether the instruments are part of a network or not. A multi-service utility (drinking water, sewer, treatment, air) uses one tool with one interface across all domains.

01 · Temporal alignment

Instruments sampling at different rates are synchronised onto a common time grid. Interpolation and realignment preserve the integrity of the original signal. Auto-detection of encoding, separator, headers, timezone, and units.

02 · Outlier detection & removal

Bad values are identified and temporarily removed from analyses. Simple outliers detected by REED can be inherited by BLEAK for deeper, domain-specific processing.

03 · Small gap filling

Short interruptions — an instrument offline for half an hour, a transmission failure — are covered with interpolations, ensuring continuity in the data flow.

04 · Unit conversion & harmonisation

Heterogeneous units of measurement and different naming conventions are normalised automatically. Any source in, one format out.

Get REED's datasheet

NEWT

Walks the pipes.

Hydraulic modelling with AI agent

NEWT combines the US EPA EPANET 2.3 engine — the most validated hydraulic calculation engine in the world, trusted for over thirty years — with a conversational interface that makes hydraulic modelling concretely accessible. The agent does not compute pressures or invent flow-rates: it orchestrates, selecting which tool to call, setting parameters, and asking for confirmation before any destructive operation.

What NEWT does

Natural language queries

"Why is pressure at J-104 dropping below 20 m at night?" The agent selects the diagnostic tool, isolates the affected DMA, proposes hypotheses (e.g. valve V-31 closed, oversized demand at node J-105). EPANET 2.3 runs the simulation and returns a verified, deterministic result. The professional reads, confirms or rejects — every step reversible.

Full GIS ingestion pipeline

Reads MIKE+, WaterGEMS, EPANET .inp, ESRI shapefiles, AutoCAD DXF, Excel. GIS preprocessing: filtering, defragmentation, connectivity repair, valve placement — guided by the AI agent, with explicit confirmation at every step.

Conversational calibration

IWA water balance, ILI, Minimum Night Flow analysis — all with a full audit trail. Undo, redo, session recovery. Professionals experiment without fear of damaging the dataset.

Native topology navigation

Navigate upstream, downstream, by district — combining spatial and temporal filters in a single conversational request.

Why it works

The engine computes. The AI reasons.

Every number comes from EPANET 2.3 — audited, deterministic engineering code. The agent chooses what to run; the physics is executed by trusted devices. No hallucinated pressures, no invented flow-rates.

No black box

The agent's reasoning is explicit and verifiable. Every action is backed by a real engineering operation. The full history is visible, searchable, reversible. The professional is always in control.

Methodology as product

Competitors expose a UI and hope the user knows what to do. NEWT exposes the methodology itself — the same approach FORE has applied across 4,500+ km of modelled Italian infrastructure.

Backbone validated on production deployments

Tested on operational networks across Italy. Not on simplified academic datasets.

Get NEWT's datasheet

BLEAK

Sees through dark water.

Sewer Data Intelligence Layer

Sewer measurements are the hardest in the field. Flow, level, and quality sensors work at their physical limits; flow-rates span two to three orders of magnitude between dry and wet weather. Raw sewer data is not simply noisy — it is systematically distorted for reasons that depend on sensor type, installation conditions, season, weather conditions, and network morphology.

Generic anomaly detection flags an outlier and stops. BLEAK names the cause. The agent orchestrates an ensemble model, identifying the right tools for each diagnostic step — it explains the cause, never just the symptom. Physics first.

What BLEAK does

Regime classification at every timestep

Classifies every timestep: dry weather, rainfall event, wet weather, pumping, overflow. Anomaly taxonomy with physical signatures: covered probe, level drift, backwater, sediment buildup.

Signal reconstruction under mass conservation

The ensemble model reconstructs anomalous periods adapting to context — dry vs. rainfall, sensor type, position in the network. A sensor in maintenance becomes a reconstructed signal, not a gap.

Continuous mass-balance check

At every node, with sensor-level fault localisation. Infiltration & inflow report, flooding proximity report — ready for hydraulic models, Early Warning Systems, or regulatory reporting.

Why it works

Recognises edge cases others miss

Dirty probe, drifting sensor, sensor below physical reading threshold, backwater masking a real event. These are the cases that fool generic anomaly detection. BLEAK was built on them.

Physics first, statistics second

Signal reconstruction respects mass conservation constraints at every node. The result is not a smoothed curve — it is an engineering-grade signal you can stake decisions on.

Continuity of service

A sensor in maintenance becomes a reconstructed signal, not a gap in the record. Operational visibility is maintained regardless of instrument availability.

Get BLEAK's datasheet