Stop Losing Water.
Start Proving It.
Precision NRW detection with sculpted multi‑modal fusion, conformal risk control, and explainability, delivered alongside OpenAPI, observability, and enterprise deployment patterns.
- Sculpted multimodal fusion & explainability
- Per‑DMA conformal thresholds & guardrails
- Signed WORM evidence bundles
- OpenAPI, /metrics & /health baked in
Metrics pair curated datasets with cross‑DMA calibration; field performance depends on telemetry quality and ongoing threshold governance.
Telemetry mix draws from Hong Kong acoustics, Yorkshire & Wessex hydraulic data, and BattLeDIM simulation. Synthetic augmentation is governed: no raw customer data persists, and waivers are logged when bypassing quality gates.
Audit‑Ready Leak Detection
Calibrated fusion with uncertainty, smoothing and guards, and evidence bundles that stand up to scrutiny.
Per-DMA conformal guard
Recall preserved while keeping false positives under control.
Explainability stack
Grad‑CAM spectrograms, SHAP features, Hydro heatmaps.
Evidence minted
WORM manifests with optional signing and verifiers.
Operational surface
OpenAPI, /metrics, /health baked into every release.
Product
Multi‑modal detection with built‑in governance
Multimodal fusion + explainability
Acoustic CNN + hydraulic detectors and HydroGNN feed a calibrated fusion meta‑learner with quantified uncertainty. Explainability artefacts (acoustic Grad‑CAM, SHAP‑like top features, and node heatmaps when available) attach to events.
Conformal thresholds per DMA
Operate in low‑FPR regimes after per‑DMA calibration. A temporal smoother and changepoint guards stabilize alerts before event lifecycle updates. Conformal quantiles (conformal_q) and thresholds are exported to metrics.
Evidence & audit trail
WORM manifests capture scores, thresholds, timestamps, and optional location; an optional signature and a verify script are emitted to support independent audit and dispute resolution.
Observability + OpenAPI
Prometheus /metrics and /health, OTEL traces, and structured logs. Key gauges (e.g., threshold_tc, threshold_tx, conformal_q) feed dashboards. Contract‑driven APIs with OpenAPI and contract tests.
Events & Attachments
Manage lifecycle via /events and /events/{event_id}; attach heatmaps and WORM manifests under /events/{event_id}/attachments.
Packaging & Deployment
SBOMs and image signing are included in releases (e.g., Syft + Cosign). Helm/K8s manifests support cloud, on‑prem, and edge rollouts; standard deployments integrate mTLS/RBAC patterns.
Architecture
From telemetry to trusted, auditable events
Ingest & Validate
SCADA/AMI and optional acoustics pass data‑quality gates first (quality_report). Valid inputs feed acoustic spectrograms and hydraulic graph features; waivers (for emergencies) are documented.
Model & Fuse
Acoustic CRNN and HydroGNN produce calibrated probabilities and uncertainties; a meta-learner fuses them, and conformal thresholds per DMA control risk.
Explain & Evidence
Grad‑CAM/SHAP‑like artefacts, node heatmaps (when inputs present), thresholds and uncertainties are attached to events; WORM evidence bundles on close.
Localize (when available)
HydroGNN heatmaps and TDOA‑based localization are provided when the request includes sensor geometry/arrivals and graph inputs are available.
Governance. Contract‑driven changes ship with updated openapi.yaml, schemas, and DB migrations plus contract tests. Builds are deterministic (seeded RNGs), secrets remain external, and every ingest emits a quality_report; non‑pass statuses block downstream use.
Integration
APIs & Contracts
| Method | Path | Description |
|---|---|---|
| POST | /predict_frame | Run calibrated inference on a single frame or window (manages event lifecycle). |
| POST | /explain_frame | Return acoustic Grad‑CAM and SHAP‑like features; HydroGNN maps via predict when available. |
| GET | /events | List recent events. |
| GET | /events/{event_id} | Fetch a specific event. |
| GET | /events/{event_id}/attachments | Evidence bundles, heatmaps, and attachments. |
| GET | /explain/{event_id} | Fetch explanation bundle for an event. |
| GET | /metrics | Prometheus scrape endpoint. |
| GET | /health | Service health for probes. |
| GET | /openapi.json | OpenAPI contract. |
| GET | /admin/thresholds | List thresholds & conformal quantiles (Admin). |
| GET | /admin/thresholds/{dma} | Get thresholds for a DMA (Admin). |
| PUT | /admin/thresholds | Update thresholds (Admin). |
| POST | /admin/reload_twin | Hot‑reload the digital twin graph (Admin). |
Example: /predict_frame
{
"dma": "dma_A",
"ts": "2025-01-01T00:00:00Z",
"spectrogram": [[0.1, 0.2], [0.3, 0.4]],
"scada_window": [
{"ts": "2025-01-01T00:00:00Z", "inlet_flow": 20.5, "press_N1": 3.1, "quality": "GOOD"}
],
"arrivals": [0.0, 0.012, 0.021],
"sensors": [[-1.23, 53.80], [-1.24, 53.81], [-1.22, 53.79]]
}{
"decision": "LEAK",
"q": 0.94,
"decision_set": ["LEAK"],
"event": {"id": "evt123", "status": "OPEN", "type": "LEAK", "q": 0.94},
"location": {"lon": -1.23, "lat": 53.80, "conf": 0.78, "valves": ["V-12", "V-13"]},
"scores": {
"acoustic_prob": 0.88,
"dma_score_xgb": 0.91,
"dma_score_gnn": 0.85,
"fused": 0.94,
"fusion_nonconformity": 0.06,
"conformal_quantile": 2.0
}
}Flow & conditions
Tap a stage to see details
Audit trail
Evidence bundle: Oct 2025 fusion run
The latest bundle packages logs, models, API contracts and charts from the 14 Oct 2025 run. Everything here is surfaced directly from the extracted artefacts so decision makers can trace performance without leaving the page.
Latest evidence bundle
The 11 Oct 2025 bundle captures a full ingest → deploy cycle with governance checks, refreshed models, and recalibrated fusion ready for executive review.
- Governance. Automated setup, ingest, feature, and deployment controls completed; final QA tuning finishes before investor circulation.
- Fusion. Multi-modal fusion holds AUROC 1.000 at a 2% false-positive rate with zero-minute detection delay.
- Evidence. WORM manifests, metrics, and audit artefacts package every stage for regulator-ready walk-throughs.
Key metrics at 2% false-positive rate
Scores come straight from fusion/metrics.json and fusion/detection_delays.json. Logistic fusion blends hydraulic, logger and acoustic signals; only the acoustic feed carried an overnight delay in this run.
Fusion
AUROC 1.00
AUPRC 1.00 and recall 1.0 at 2% FPR
Detection delay 0 minutes
Threshold 0.716
Hydraulic
AUROC 0.995
AUPRC 0.9997 and recall 0.96 at 2% FPR
Detection delay 0 minutes
Logger (Wessex)
AUROC 0.985
AUPRC 0.990 and recall 0.68 at 2% FPR
Threshold 0.859
Acoustic
AUROC 0.754
AUPRC 0.878 and recall 0.42 at 2% FPR
Delay 865 minutes (overnight)
Model performance highlights
Hydraulic, logger, and fused models remain inside customer SLA bands, with the latest fusion pass delivering perfect AUROC and recall at operating thresholds.
- Fusion. AUROC 1.000, AUPRC 1.000, recall 1.0 @ 2% FPR, operating threshold 0.716.
- Hydraulic. AUROC 0.995 with zero detection delay across the monitored DMA set.
- Yorkshire logger. AUROC 1.000 with full recall at the 2% FPR target.
- Wessex logger. AUROC 0.985 and calibrated recall 0.68 @ 2% FPR.
Request the evidence bundle for dashboards, precision-recall curves, and audit manifests aligned to these figures.
Run timeline
Automated run controls span data readiness through deployment. Continuous QA keeps the bundle live while final threshold tuning completes.
-
Setup & data readiness
Environment established with telemetry feeds verified and governance gates primed for ingest.
-
Code quality sweep
Static analysis automation flagged a styling rule scheduled to close before investor release.
-
Automated tests
Regression suite ran 284 scenarios; a fusion guardrail is being recalibrated to complete QA.
-
Ingest & features
Authentic and governed telemetry streams were ingested, cleansed, and engineered for multimodal training.
-
Model refresh
Hydraulic and logger models retrained with documented adjustments to maintain stability across networks.
-
Fusion calibration
Logistic fusion retuned across modalities, preserving AUROC 1.000 and operational thresholds.
-
API validation
Smoke deployment confirmed health responses, contract checks, and attachment workflows.
-
Patch assurance
Final container patch verified runtime parity ahead of customer rollout.
Charts direct from the artefacts
These high resolution exports live under evidence_bundle/charts/ so what you see here is exactly what auditors download.
API smoke check
Health and /predict responses are streamed directly from api_health.json and api_predict.json so operators can rehearse integrations.
{
"detail": {
"code": "scada_contract",
"message": "ok",
"details": {
"ok": true,
"issues": []
},
"trace_id": "local-sample"
}
}
Trace local-sample shows the contract guard responding with scada_contract. Use the OpenAPI snapshot for the full schema.
Evidence
Benchmarks & caveats (curated datasets)
- Hydraulic/logger models show AUROC ≈ 0.98–0.999 at low false-positive rates on curated datasets. The acoustic CNN is earlier-stage (≈0.75 AUROC on HK set).
- Field performance varies with telemetry quality and per-DMA calibration. We operate in low‑FPR regimes after calibration, enforced via a conformal guard.
- Detection latency depends on cadence: continuous telemetry enables near-real-time detection; acoustic nightly programs are typically overnight; satellite scans periodic.
- Datasets referenced: Hong Kong acoustics; Yorkshire & Wessex loggers/hydraulic; BattLeDIM simulation (for stress & coverage). Synthetic augmentation follows documented governance.
Figures shown are illustrative unless tied to current run artefacts; use our OpenAPI & evidence bundles to audit live performance.
Operations
Deployment, security & observability
Cloud / On‑prem / Edge
Deploy as containers with Helm/K8s. SBOMs and image signing are included in the release process; signing requires a configured signer/HSM or software key.
Observability
Prometheus metrics, health endpoints, traces and structured logs for end‑to‑end visibility and SLOs.
Evidence & Compliance
WORM evidence bundles with optional signing and an on‑close attachment flow; audit verifiers provided.
Economics
ROI examples (assumptions disclosed)
Mid-case
Illustrative ~5× ROI with continuous monitoring and calibrated thresholds; depends on leak incidence and water value.
High-scarcity
Higher ROI in regions with expensive water or strict NRW targets; energy/carbon savings add to value.
Case studies
External case studies show large savings with event analytics and satellite-augmented programs; our platform integrates these strengths.
Investors
Investor brief
NRW reduction is a board‑level priority for many utilities. Our platform focuses on audit‑ready detection built on authentic telemetry with governed synthetic augmentation, contract‑driven releases, and evidence that stands up to scrutiny.
Why now
Utilities face tighter NRW targets, water scarcity pressures, and audit requirements. Modern telemetry and explainable AI enable measurable, reportable improvements.
Product moat
Conformal thresholds per DMA, WORM evidence bundles (optional signing), OpenAPI + observability, and strong governance (quality gates, seeded builds, signed artefacts).
Model & GTM
SaaS per‑DMA subscription with optional services for onboarding and integration. Land‑and‑expand via pilot → region → utility‑wide rollouts; channel partners where helpful.
Economics
ROI comes from shorter leak duration, fewer truck rolls, energy/carbon savings, and avoided penalties. See ROI examples; field outcomes depend on telemetry quality and calibration.
Compliance & trust
Evidence bundles, signed artefacts, and contract tests support audits and regulated reporting. Observability and SLOs de‑risk operations.
Risks & mitigations
Data quality and integration friction are addressed via ingest gates, conformal guards, shadow deployments, and acceptance runs before cutover.
What we’re seeking
Pilot partners and design partners to validate in diverse DMAs, plus investor conversations aligned with an evidence‑first rollout plan. For materials (decks, acceptance run outputs), reach out via the contact details below.
Roadmap
What’s coming
Theft detection
Enhanced scenarios and classifiers (synthetic generators already include theft cases).
Satellite fusion UI
UI overlays and fusion controls for periodic satellite layers.
Leak sizing
Expose leak sizing estimates in operator workflows with uncertainties.
Certifications
Security attestations and programme‑level M&V reports.
Contact
Get in touch
Contact Us
We'd love to hear from you. For engagements, partnerships, or media enquiries, use the details below.
Maqbool Ahmed Inamdar
Founder and Chief Executive Officer
MSc Artificial Intelligence, University of Essex
London, United Kingdom
Haseena Inamdar
Cofounder and Business Advisor
B Tech Electronics and Communication Engineering, Visvesvaraya Technological University (VTU)
Bengaluru, India
Ahesan Ali
Cofounder & Chief Operating Officer
MSc Data Science, University of Essex
Colchester, United Kingdom
Khadar R J
Business Admin
B Tech Computer Science and Engineering, Visvesvaraya Technological University (VTU)
Bengaluru, India