Mobile connectivity

I worked on Vodafone's IoT platform at the intersection of customer cost confidence, machine learning, and developer experience.

For many IoT customers, spend volatility isn't an edge case, it's an operational reality driven by rollout phases, device misconfiguration, roaming exposure, firmware changes, and shifting usage patterns across fleets. The result was familiar: teams could see what happened after the billing period closed, but they lacked a reliable way to anticipate the bill, set guardrails, and catch abnormal behaviour early enough to act.

That gap created three business-critical problems:

I helped shape a machine-learning-driven predictive analytics platform that reframed cost management from reactive reporting into proactive control. The product goal was simple: give customers an answer to "What will we spend?", the tools to "Keep it within bounds," and the signals to "Investigate what changed, now," without requiring them to become data scientists.

Forecasting: from hindsight to forward view

I designed forecasting around how customers plan and govern IoT programs: forecasted spend for the current and upcoming billing periods, not just a rolling chart, so it could be used in budget conversations; breakdowns that matched customer mental models (fleet/group/program, geography, device/SIM cohorts, plan/tariff), so "why is it changing?" was answerable without jumping between tools; clear communication of uncertainty so the forecast didn't present a false sense of precision, with expectations for when confidence is high vs. when patterns are shifting; comparison to historical baselines to make trend movement interpretable ("this week is trending above last month's comparable window"). The design work wasn't "hide the model." It was "make the model usable": transparent inputs, understandable outputs, and decision-ready context so customers could act with confidence.

Spend quotas: turning budgets into guardrails

Forecasting is insight; quotas are control. I worked on enabling customers to create spend quotas that reflected real governance structures: quotas scoped to the level customers actually manage (account/program/group/cohort), with flexibility to match how budgets are allocated; thresholds that created early-warning moments (approaching limit vs. exceeded), so teams could intervene before the invoice; a deliberate "what happens next" path (investigate top contributors, adjust quotas with governance, and set watch conditions for high-risk segments) so quotas became part of an operating rhythm, not a set-and-forget screen; traceability of changes (who set what, when) to support internal accountability between finance and operations.

Anomaly detection: catching abnormal spend before it becomes a surprise

I designed anomaly detection to prioritise signal quality and speed-to-diagnosis: detection tuned to common IoT failure modes (sudden spikes, gradual drift, and cohort outliers where one segment behaves unlike peers); explanations written for operators ("what changed, when it started, and how large the deviation is"), not just statistical labels; investigation flows that reduced time-to-root-cause (quick slicing by segment and time window, baseline comparisons, and direct links to the underlying usage patterns driving the alert); alerting that fit real workflows, so anomalies could be monitored and actioned, not simply observed.

Developer documentation: making the platform truly self-serve

Because many customers operationalise cost management through automation, I also worked on developer documentation to reduce time-to-first-success and ongoing integration risk: a clean structure (quickstart, guided workflows, API reference, troubleshooting) so teams could choose "fast" or "deep" without getting lost; end-to-end examples for common jobs-to-be-done (retrieving usage/spend signals, setting quotas, subscribing to alerts/events, and integrating anomaly signals into existing systems); consistent vocabulary between the UI and the API so "quota," "threshold," and "anomaly" meant the same thing everywhere; practical troubleshooting content based on real implementation failure modes, reducing reliance on support as the default path.

This work shifted the customer experience from retrospective billing analysis to proactive cost management: forecast what's coming, set boundaries, and catch abnormal behaviour early. That change matters because it unlocks organisational alignment: finance can plan, operations can respond quickly, and program owners can scale deployments without fear of uncontrolled spend.

It also reduced avoidable operational drag. When anomalies are detected early and attribution is clearer, teams spend less time in escalations and more time taking corrective action, often before a billing cycle closes.

Finally, the documentation work compounded product adoption by making the capabilities implementable in customer environments. Clear, task-oriented developer content shortens onboarding, improves integration reliability, and makes the platform stickier because cost controls and alerts become embedded into daily operations rather than living in a dashboard.