Turning fragmented data into trusted, AI-enabled decisions — at enterprise scale.

In early 2025, the IT support function was running lean, fewer support agents after company layoffs, while reskilling toward an AI-enabled support model, and it showed:

• Case spikes were handled reactively
• Executive leadership was often the last to know, hearing about problems only once they'd already escalated

This led me to develop a two-tier tripwire system based on standard deviations over six months of case trends, tuned tighter for operations and looser for executives, then led the development of an AI Assistant to summarise the essential pieces of information using prompt engineering techniques such as question refinement, cognitive verifier and few-shot examples.

We then surfaced these tripwire notifications up in IM chat – meeting where the user already is - and have it answer questions straight from a verified, business-ready SQL views.

Case-spike analysis went from half a day to near-instant, and the surprises between operations and executive leadership simply stopped.

Cisco's laptop refresh programme had always run on age, not evidence: three years and you're up for refresh, regardless of whether a device was struggling or still going strong.

That left asset management fragmented, budget handling inefficient, and refresh timing badly out of step with what users actually needed.

Rather than reach for the obvious generative AI answer, I ran a genuine build-versus-buy evaluation, supervised learning, GenAI, pure statistical thresholds, before landing on unsupervised anomaly detection:

• No training data to maintain
• None of GenAI's hallucination risk
• A model that scaled naturally as the device landscape changed

Ranking real performance from CPU, RAM, memory pressure and disk telemetry, combined with app usage-based personas, it became the foundation for AI-enabled laptop performance scoring, an intelligent refresh decision engine and predictive stock forecasting, turning refresh decisions evidence-based and unlocking millions of dollars in potential savings.

When RAM costs rose and new supply tariffs pushed Apple's M5 MacBook Pro pricing up, every future Mac refresh using those models was set to get materially more expensive, a cost increase nobody had budgeted for and nobody had proven was actually necessary.

So I had the team quantify CPU and RAM performance across the entire Mac and Windows fleet, benchmarked it against the new M5 MacBook Air, and forecast the cost avoidance of standardising on the Air instead of the Pro.

The evidence made the decision straightforward:

• 1.5x the performance of the M4 generation, double the M3
• 23% less than the M5 14 inch and under half the cost of the 16 inch
• Comfortably meeting the needs of over 87% of Mac users globally

Enough to make the Air Cisco's new standard on the spot.

As AI tools became part of daily work, the same signal kept showing up in laptop performance and app-usage data: a growing group of engineers and AI developers whose devices simply couldn't keep pace with what their jobs now demanded.

I used Python to analyse historical device performance, isolated the users with anomalous CPU and RAM growth, and cross-checked those spikes against the applications actually driving them, developer tools and IDEs with AI add-ons among them, before building the case for Finance with Power BI visualisations.

I didn't expect it to be approved on the spot, and it wasn't; the goal was to put verified evidence on Finance's radar early and keep advocating with that same evidence into the next financial year, rather than wait for the problem to force a rushed decision.

A third of all data requests were simple enough that no one should have needed an analyst, yet they cost 512 hours of stakeholder wait time and $58K in lost productivity a year, pulling analysts away from the work that actually needed them.

I introduced ThoughtSpot as a self-service layer instead, built entirely on the same verified, governed datasets used for executive KPI dashboards, so speed never came at the cost of trust, then added Spotter as a natural-language GPT layer on top for forecasting and correlation analysis, with new users automatically enrolled in onboarding material from day one.

Adoption grew from 46 to 196 users (+326%), monthly queries from 8.1K to 19.2K (+137%), while the data team's own request workload fell 33%, freeing an estimated 738 hours and $68.9K a year for higher-value work.

ThoughtSpot ran as an internal shared service through Cisco's Data & Analytics Office, and as our user base grew across three other departments, that arrangement stopped making sense:

• We were being cross-charged at a higher rate than dealing with the vendor directly
• Administration depended on an internal support team with limited capacity that was slow to respond
• The DnA Office wasn't upgrading to the latest ThoughtSpot version, so the newest technical and AI capabilities never reached us

Rather than keep absorbing that cost and lag, I partnered with other organisational data leads and went direct to ThoughtSpot's account management team to stand up our own instance.

Negotiating the contract terms and cost ourselves and agreeing to share that cost across the departments now using it, trading a shared service for one we owned and controlled outright.

That direct relationship put full administration and the latest capabilities in our own hands, and gave us the leverage to build the commercial case for a capacity uplift:

• Switching to a $70K/250M-row option delivered 2.86x better value than the $40K/50M-row option
• 65% lower cost per million rows
• 5x the capacity for only 1.75x the cost