Our work.

Cloudpital is our own integrated, locally compliant health record platform. Its AI layer runs the patient journey end to end and engages patients over WhatsApp at every step, so people arrive when the clinician is ready for them instead of waiting in a room.

• Integrated record covering registration, clinical documentation, pharmacy, claims and billing
• Patient engagement over WhatsApp across booking, preparation, arrival and follow-up
• Scheduling driven by live clinic state rather than a static appointment grid

Facility, workforce and service data arrived from hundreds of sites in different shapes. We standardised the reporting model, deployed the analytics platform and trained the authority's own teams to extend it.

• Standard reporting model applied consistently across facility types
• Executive, regional and facility views built from one certified layer
• Internal analyst enablement so the estate grows without external dependency

During Hajj and Umrah seasons, health operations run at extraordinary density for a fixed window. We deliver the monitoring layer that tracks service load, capacity and incidents across sites so command teams can move resources while the season is still running.

• Site-level load and capacity tracked through the operating window
• Incident and escalation reporting aligned to command structures
• After-season review built from the same data the command centre used live

Arrival forecasting and a clinician-facing triage support layer, built on the group's own historical presentations and staffed rosters, with clinicians retaining every decision.

• Hourly arrival forecasts feeding roster and bed planning
• Triage support surfaced as a suggestion with reasons, never as an instruction
• Model performance reviewed by the clinical governance committee

A governed assistant that drafts clinical notes and suggests coding from the consultation record, with the clinician approving every entry before it enters the record.

• Nothing enters the record without explicit clinician approval
• Coding suggestions traced to the note text that justified them
• Patient data stays inside the group's governed environment

A ministry-wide reporting platform replacing directorate-by-directorate compilation, with a shared measure dictionary and role-based access from leadership down to operating units.

• One measure dictionary agreed across directorates
• Role-based access with sensitive material scoped to cleared audiences
• Scheduled refresh with monitored data quality

Permit, inspection and service-request data unified across municipalities, with backlog, cycle time and service-level views that let regional teams see where a queue is forming before it becomes a complaint.

• Comparable cycle-time definitions across municipalities of very different size
• Backlog ageing visible by service, region and responsible unit
• Service-level breaches surfaced to the owning team, not just to the centre

Agency data centralised into one governed platform, with an executive view that connects service delivery, spend and workforce in a single reading rather than three separate reports.

• Agency onboarding pattern reused rather than rebuilt per entity
• Executive view assembled from certified measures with named owners
• Access governed by entity and role

Case, referral and disclosure data brought into a governed analytical layer with executive dashboards, so oversight leadership can see workload, ageing and pattern signals across the portfolio.

• Sensitive material handled under strict access scoping and full audit logging
• Case ageing and workload visible by unit and stage
• Analytical signals presented as leads for review, never as conclusions

Forecasting of incoming case volume by channel and service, paired with a triage layer that routes straightforward cases to a fast path and complex ones to specialist teams.

• Routing rules published and reviewable, with appeal paths preserved
• Forecast accuracy monitored and reported to the service owner
• No automated adverse decision, routing only, decisions stay with caseworkers

We combined the student information system, survey platforms and departmental sources into a single university data warehouse, then built the Tableau reporting estate over it, admissions, progression, teaching load and student experience read from one model.

• Student information system, survey data and departmental sources integrated into one warehouse
• Reporting estate covering admissions, progression, teaching load and student experience
• University KPI reporting produced from certified measures rather than ad-hoc extracts

Student attendance capture rolled out to 1,200 teaching staff on managed tablets, with geofencing so attendance can only be recorded inside campus boundaries, and reporting that closes the loop back to programme administration.

• Attendance recordable only inside the defined campus perimeter
• Managed device fleet with controlled configuration across faculties
• Attendance feeding programme reporting rather than sitting in a separate app

Early-signal modelling on engagement and assessment data, delivered to academic advisors as a prioritised outreach list with the reasons behind each flag.

• Flags carry reasons, so an advisor starts the conversation informed
• Outreach outcomes recorded and fed back into model review
• Students are never scored in a way that affects assessment

Optimisation of timetabling against room capacity, staff availability and student pathways, replacing a manual scheduling cycle that took weeks each term.

• Constraints declared explicitly, so a rejected timetable explains itself
• Scenario comparison before publication
• Schedulers keep final control of the published plan

Line, test and returns data unified across plants so a quality signal in the field can be traced back to the shift, line and component batch that produced it.

• Field returns joined to production and test records at batch level
• Comparable line performance measures across plants
• Deviation investigations start from evidence rather than from a workshop

A shared performance model across plants, output, downtime reasons, changeover and yield, replacing plant-specific spreadsheets that could not be compared.

• One downtime reason taxonomy used by every plant
• Shift-level reporting available to supervisors, not only to the centre
• Losses quantified in output terms leadership already tracks

Condition models on the assets whose failure stops the line, integrated into the maintenance planning system so a prediction becomes a scheduled job rather than an alert nobody owns.

• Coverage limited to assets with a real production consequence
• Predictions raised as maintenance work orders with an owner
• Model performance reviewed against actual failures each quarter

Inline visual inspection that flags defects at the station where they occur, with operators confirming or overriding and every judgement retained for model review.

• Inspection at the originating station rather than at the end of the line
• Operator confirmation retained as labelled training evidence
• Rejection thresholds owned by quality, not by the model

Meter, network and billing data reconciled into one governed foundation, making the difference between energy delivered and energy billed explainable feeder by feeder.

• Delivered-versus-billed reconciliation at feeder level
• Data quality issues in metering surfaced as work, not written off as loss
• Field investigation prioritised by expected recovery

Short-horizon forecasting of generation and demand feeding dispatch and trading decisions, with forecast error tracked against its financial consequence rather than as an abstract accuracy score.

• Forecast error reported in cost terms the trading desk uses
• Weather and asset-availability inputs versioned and auditable
• Fallback behaviour defined for input outages

Incident, near-miss and inspection data consolidated across sites and contractors into one reporting foundation with consistent severity classification.

• Contractor and operator incidents classified on the same scale
• Leading indicators reported alongside lagging ones
• Site-level ownership of every open action

Warehouse, transport and customer-promise data joined into one operating view so exceptions surface while the shipment can still be recovered.

• Exceptions routed to the team that can act, with a time-to-act target
• Promise dates modelled from actual operating performance
• Customer-facing status derived from the same data as internal operations

Optimisation of routing and load consolidation against real network constraints, with planners comparing scenarios before committing a plan.

• Constraints reflect real depot, vehicle and driver-hour limits
• Planner override retained and analysed rather than blocked
• Savings measured against the plan that would otherwise have run

Terminal operating data modelled into berth, yard and gate performance measures, exposing where dwell accumulates and which moves create it.

• Dwell decomposed by stage rather than reported as one figure
• Gate, yard and berth measured on one clock
• Shift teams see their own performance in near real time

Large-scale reporting and applied AI work across merchandising, supply and store operations, with forecasting and replenishment signals delivered into the systems buyers and store teams already use.

• Forecast and replenishment signals delivered inside existing buying workflows
• Store-level measures consistent with enterprise reporting definitions
• Model outputs reviewed against realised sell-through, not just accuracy

Store-and-item level forecasting tuned to local demand patterns, seasons and religious calendars, feeding daily ordering for fresh categories.

• Forecasts respect local calendars and event demand, not a generic seasonality curve
• Order recommendations adjustable by store managers with the reason recorded
• Waste and availability tracked together so one is not improved at the other's cost

Assortment planning and markdown timing supported by models on sell-through, size curves and store clustering, with buyers retaining the final call.

• Store clusters derived from behaviour rather than geography alone
• Markdown recommendations explain the expected trade-off
• Buyer overrides tracked and used to improve the model

An enterprise BI platform for commercial and network reporting, replacing per-market reporting stacks with one governed model that markets extend rather than duplicate.

• One subscriber and revenue definition applied across markets
• Market-specific extensions built on the shared model, not beside it
• Refresh, quality and usage monitored centrally

Modernisation of the commercial reporting estate, consolidating legacy extracts into a governed model and rebuilding the reports the business actually runs on.

• Legacy extracts retired rather than wrapped
• Reports rebuilt against certified measures with named owners
• Usage measured so unused reports are decommissioned

Churn models joined to the retention offer engine, so the prediction lands as a specific offer to a specific customer through the channel they use, with holdout groups proving the effect.

• Holdout groups maintained so the measured effect is real
• Offer economics enforced, no retention at a negative margin
• Contact frequency governed across campaigns

Alarm, performance and customer-impact data correlated into incident candidates, so engineers start from a probable cause rather than from a wall of alarms.

• Alarm storms collapsed into ranked incident candidates
• Customer impact attached to every candidate
• Automation limited to triage; remediation stays with engineers

A group-wide data quality programme across property systems, the reporting layer built on top of it, and AI assistants that answer guest and staff questions from records the group can trust.

• Guest identity resolved across properties and booking channels
• Quality rules enforced at intake, not corrected downstream
• Assistants answer only from governed records, with escalation to staff

Forecasting of demand by segment and channel feeding pricing and inventory decisions, with revenue managers comparing scenarios before publishing rates.

• Event, season and calendar effects modelled explicitly
• Recommendations explain the expected trade-off between rate and occupancy
• Revenue managers retain publication control

A governed assistant handling booking changes, property questions and service requests in the guest's own language, with clean escalation to staff whenever judgement is required.

• Answers grounded in property records rather than generated from general knowledge
• Escalation to a human preserves the full conversation context
• Guest data handled under the group's own governance rules

Product, billing and support data unified so product, sales and finance work from the same definition of usage, adoption and account health.

• Event schema governed and versioned rather than added ad hoc
• Account health composed of measures each team recognises
• Definitions published with owners

A grounded assistant over documentation and prior resolved tickets, plus clustering that turns repeat contacts into prioritised engineering signals.

• Answers cite the document or ticket they came from
• Unresolved contacts escalate with full context
• Repeat themes reported to engineering with volume and cost attached

Project delivery, sales and cost data joined into one portfolio view, so schedule slippage and its commercial consequence are visible in the same reading.

• Delivery milestones and sales pipeline reconciled per project
• Cost variance attributed to a cause, not just reported
• One portfolio review pack generated from the platform

Building management, occupancy and lease data combined so energy and service cost is read against actual occupancy rather than floor area.

• Cost normalised by occupancy, making buildings comparable
• Anomalies raised as maintenance work with an owner
• Reporting aligned to lease service-charge cycles

Consolidated 14 legacy retail-banking marts into a single on-premise lakehouse (Iceberg + Trino) feeding a sovereign voice-agent platform in English across the region. Front-office bankers now query 2.3PB of customer data in plain English through a governed natural-language-to-SQL layer, every answer cited to a certified metric, every voice interaction fully on-prem.

• 2.3PB Iceberg lakehouse on-prem, zero customer data leaves the bank's perimeter
• Natural-language-to-SQL layer over 380 certified metrics, used by 4,200 branch staff
• Voice-agent deployment live in 9 weeks across the region, integrated with Genesys and Salesforce FSC

Unified Epic, Cerner and 9 ancillary systems into a HIPAA-audited Databricks lakehouse (1.8B clinical events) powering a multi-agent orchestrator for scheduling, refills and post-discharge follow-up. Clinicians and ops leaders query population-health dashboards in natural language, no-show rate fell from 19% to 7% and 140 clinical FTE hours a week were returned to bedside care.

• 1.8B clinical events unified in a SOC 2 Type II + HIPAA-audited lakehouse
• Multi-agent orchestrator handling 3.1M patient interactions per year
• Clinician-facing dashboards with plain-English querying over 240 certified measures

Migrated network, billing and CRM feeds into a Databricks Mosaic AI lakehouse (1.1PB) running daily churn and CLV scoring on 6.8M subscribers. Retention teams work a natural-language analytics workspace that turns 'show me high-value customers likely to churn in Sydney this month' into governed SQL against certified features, campaign ROI up 4.2×.

• 1.1PB streaming lakehouse in-region on Azure Sydney
• 6.8M subscribers scored daily; real-time triggers to Genesys and the operator's mobile app
• Natural-language analytics workspace adopted by 380 retention analysts across the region

Deployed Collibra + Microsoft Purview across a 240-dataset Azure Data Lake, automated personal-data classification and end-to-end lineage from source to Power BI dashboard. Policy analysts now ask questions in plain English via a governed NL→SQL layer that refuses queries against uncertified fields, the first Danish GDPR + Datatilsynet audit passed with zero remediation.

• Collibra + Microsoft Purview rollout in 14 weeks across 240 datasets
• Automated personal-data discovery with end-to-end lineage to every Power BI report
• NL→SQL layer restricted to certified, RACI-owned fields across 9 policy domains

Consolidated SAP S/4HANA, MES and IoT sensor streams from 14 plants into a Snowflake + dbt + Fivetran lakehouse, retiring 7 legacy data marts. Shift managers now open a plant-effectiveness dashboard and ask 'why did line 3 in Espoo drop yield last night?' in English; the semantic layer resolves the answer against certified plant KPIs in seconds.

• Snowflake + dbt + Fivetran on Azure Nordics unifying SAP S/4HANA, MES and IoT across 14 plants
• Production dashboards live in 11 weeks (down from a 9-month legacy roadmap)
• English natural-language querying over a governed semantic layer

A Databricks lakehouse of 11 years of motor + home claims feeds Azure Document Intelligence and a hybrid rules-plus-ML decisioning engine. Claims leaders use a natural-language cockpit, 'show me open bodily-injury reserves above A$50k in Queensland', instead of raising a BI ticket. Median cycle time fell from 9 days to 32 hours.

• 11-year claims lakehouse on Databricks with governed feature store
• Document AI + custom underwriting rules engine settle 68% of eligible claims straight-through
• Underwriter NL cockpit replaces 90% of ad-hoc BI ticket volume

Grid telemetry, weather and renewables generation from 6 markets stream into an open lakehouse (Iceberg on S3) with probabilistic 15-minute-horizon forecasts. Traders query positions, imbalance risk and asset performance through an NL analytics layer over a certified semantic model, no more Excel exports of dispatch data.

• Open Iceberg lakehouse ingesting 4 years of grid telemetry across the Nordic markets
• 15-minute probabilistic forecasts delivered to trading desks and TSO APIs
• Certified semantic layer powering trader dashboards + NL analytics

POS, e-commerce, inventory and workforce data from 1,400 stores land in a Microsoft Fabric OneLake with a Power BI executive cockpit refreshed every 15 minutes. Store managers, regional VPs and the CEO ask the same dashboard questions in English, the NL-to-SQL layer answers off a single certified metrics catalog. Same-store revenue up 14% in six months.

• Microsoft Fabric OneLake + Power BI cockpit across 1,400 stores, 15-min refresh
• Tri-lingual EN natural-language querying on a certified 210-metric catalog
• Store manager mobile dashboards driving daily labour + markdown decisions

Consolidated 6 legacy AML systems into a Snowflake lakehouse of 4.2B transactions with a graph-plus-ML fraud model and a supervised co-pilot that drafts SAR narratives, explains model scores and answers analyst questions in English via NL→SQL against certified investigation features.

• 4.2B transactions unified in a Snowflake fraud lakehouse with graph features
• SAR narrative drafting cut writing time from 45 to 9 minutes on average
• English NL analytics workspace adopted by 620 investigators

12 months of telematics, orders and depot data land in a Delta Lake feeding a nightly route + load solver. Planners use a English NL analytics layer to interrogate driver-hour compliance, empty-leg kilometres and lane profitability, no BI engineer in the loop.

• Delta Lake with 12 months of telematics + order data across the fleet
• Nightly optimisation solver retrained on latest lane and driver constraints
• English NL analytics dashboards for depot planners, no BI ticket required

A sovereign lakehouse consolidates 380 national-program KPIs with a bilingual Arabic/English NL analytics workspace. Every executive answer carries policy-grade citations and a traceable lineage back to the source indicator, reducing the time from question to Cabinet-ready briefing pack by 82%.

• Sovereign lakehouse holding 380 national-program KPIs with certified lineage
• Bilingual EN/AR NL-to-SQL layer with policy-grade citations on every answer
• Executive dashboards traceable line-by-line to source policy or indicator

A challenger credit model on a Feast feature store (420 features, 8 years of history) generates regulator-ready explanations for every declined applicant. Risk officers query approvals, expected loss and vintage curves in plain English via a governed analytics layer, 9-point approval lift at the same expected loss.

• Feast feature store: 420 governed credit features across 8 years of history
• Model card + reason codes reviewed by second-line risk on every decision
• Adverse-action letters generated automatically in 3 languages

PMS, channel-manager, competitor-rate scrape and market-demand signals from 60+ properties across Middle East and Nordic properties land in a Snowflake revenue lakehouse. A nightly pricing model recommends rates; revenue managers accept, override or ask 'why is Dubai ADR softening next weekend?' in a natural-language cockpit, accepted 71% of the time without change.

• Snowflake revenue lakehouse across 60+ properties in 14 countries
• Nightly pricing model with hold-out control group across the portfolio
• Revenue-manager NL cockpit answering pricing 'why' questions on demand

A guest-360 lakehouse stitches bookings, on-property IoT, dining, spa and excursion signals into a real-time predictive model that personalises itineraries and pre-empts service issues. Operations leaders query 'which arriving VIPs have unmet dietary requests?' in natural language, 63% of predicted service incidents resolved before the guest complained.

• Real-time guest-360 lakehouse ingesting bookings, IoT, dining and excursion signals
• Personalised itineraries for 100% of arriving guests before check-in
• Operator NL analytics workspace over certified guest + revenue KPIs

An always-on voice + chat AI concierge handles reservations, restaurant bookings and activity requests in English and Arabic, backed by a guest lakehouse that keeps PMS, POS and booking-engine data in one certified model. Duty managers use an NL dashboard to see live sentiment, deflection rates and open recovery cases.

• 9-language voice + chat integrated with PMS, POS and F&B booking engines
• 42% deflection of concierge phone volume in the first 90 days
• Duty-manager NL dashboard over live guest + concierge KPIs

Unified guest identity across 8 source systems (PMS, loyalty, spa, F&B, digital) in a governed lakehouse with consent tracking. Next-best-action surfaces in the PMS in <400ms; marketers query campaign performance in plain English against a certified 180-metric catalog. 3.2× uplift on repeat-stay conversion.

• Unified guest identity across 8 source systems with GDPR + consent tracking
• Front-office next-best-action surfaced in PMS in under 400ms
• Marketer NL analytics on a certified 180-metric guest catalog

A claims lakehouse scoring 27M outpatient claims per month combines a hybrid graph + gradient-boosting fraud model with an agentic reviewer that drafts investigator notes. Fraud managers query provider-level leakage and case status in natural language, outpatient leakage down 31% in six months.

• 27M claims/month scored with sub-200ms latency in the claims lakehouse
• Agentic reviewer drafts investigator notes with full audit trail
• Investigator NL dashboards over certified fraud + leakage KPIs

A Feast feature store serving 640 features at 45k QPS feeds a real-time decisioning engine on device telemetry, telco and merchant data. Risk officers query approval rates, cohort defaults and challenger-vs-champion performance in plain English against a certified feature catalog, approval time down from 3 days to 8 seconds median.

• Feast feature store: 640 features at 45k QPS with shadow-scoring in prod
• Real-time decisioning on device, telco and merchant signals
• Risk-team NL analytics over the certified feature catalog

SCADA, weather and drone-inspection imagery stream into an open lakehouse (1.4B events/day) that scores every transmission asset for failure risk. Dispatcher-facing dashboards link each alert to a maintenance work-order, and grid engineers query 'top 20 assets at risk in the Ostrobothnia corridor' in English.

• 1.4B sensor events/day ingested into an open Iceberg lakehouse
• Asset-level failure-risk scoring feeding dispatcher dashboards
• English NL analytics over certified grid + maintenance KPIs

Edge vision models catch defects at 40ms/frame on plant hardware; a Databricks lakehouse stitches PLC, MES and quality data across 14 sites into a single certified dashboard. Shift managers ask 'which lines dropped OEE last shift?' in natural language, no dashboard requests, no Excel.

• Databricks plant lakehouse across 14 sites with certified OEE semantic layer
• Edge vision inference at 40ms/frame on plant-floor hardware
• Shift-manager NL dashboards replacing 85% of ad-hoc plant BI requests

12k SKUs re-priced daily across 400 stores and e-commerce from a Snowflake pricing lakehouse combining elasticity models, competitor scraping and inventory constraints. Category managers approve, override or interrogate every recommendation in English via an NL cockpit over a certified pricing catalog.

• Snowflake pricing lakehouse re-scoring 12k SKUs daily across 400 stores + e-com
• Elasticity + competitor + inventory signals unified in a certified catalog
• English NL cockpit for category managers with full override audit

Document AI parses FNOL packets into a Databricks lakehouse; a decisioning workflow orchestrates fraud, coverage and reserve checks with a full EU AI Act & Nordic supervisory model-card trail. Claims leads query loss ratios and exception queues in plain English against a certified metric layer.

• Databricks claims lakehouse with EU AI Act & Nordic supervisory–compliant decision + model logs
• 68% of motor claims settled straight-through in under 6 minutes
• English NL dashboards for claims leads and reserving actuaries

TMS, WMS, telematics and customs data across the region fuse into a streaming lakehouse (46k weekly shipments). A live control-tower dashboard highlights at-risk shipments; planners ask 'which cross-border lanes are slipping this week?' and a replanning agent proposes reroutes.

• Streaming lakehouse across the region and 46k weekly shipments
• ETA accuracy within ±35 min on cross-border lanes
• Planner NL control tower with agentic reroute proposals

RAN, transport and core telemetry (1.2M metrics/sec) stream into a lakehouse feeding anomaly models that surface P1/P2 incidents minutes ahead. A runbook agent executes safe remediations; NOC engineers query incident history and self-heal rates in natural language against certified network KPIs.

• 1.2M metrics/sec across 4 network layers in a streaming lakehouse
• Anomaly models surfacing P1/P2 incidents minutes before escalation
• Self-healing runbook agent resolving 38% of triaged tickets

9.4M taxpayers scored monthly in a consolidated VAT + corporate + third-party lakehouse. Audit officers work a prioritised queue with model-explainability panels; they can also ask 'top 50 taxpayer risks in the construction sector last quarter' in the national language against certified risk features.

• Consolidated tax lakehouse scoring 9.4M taxpayers monthly with governance-approved features
• Auditor NL analytics workspace with full lineage back to filings
• Model cards and lineage evidence accepted by national audit office