Work

Problem, approach, outcome.

Engagements and projects, written up as what was actually wrong, what I did about it, and what changed. The pattern is the same every time: walk into something undefined, map the work before building anything, prove the approach on something small, then build out from there.

Case studies

Helm Wealth · 2023 → 2026

The AI Programme

4: Business units in production
0: CRM writes without human approval

Problem: A regulated wealth manager with no AI capability, no policy, and a board that needed to see governance before it would approve anything.
Approach: Designed the AI programme end to end. Wrote the AI policy. Ran a readiness audit across SharePoint and the wider Microsoft environment to see what data AI would actually touch. Implemented DLP to secure our data and ZTNA to block unauthorised AI use. Built an intake process for new use cases, including a privacy assessment before anything ships. Then shipped: Copilot and ChatGPT into daily use across the firm, and a production CRM agent for the sales team. The CRM agent is the piece I'm most proud of — designed and built on an MCP server that connects to Dynamics 365, running inside Microsoft Teams. The sales team uses it daily to prepare for client meetings, generate weekly account summaries, capture new leads, and add follow-up tasks. Nothing writes back to the CRM without human approval. The risk committee signed off on it before it went live. Alongside it, agents that streamline investment research and content creation.
Outcome: AI is in production across four business units. The governance practice means new use cases go through a defined pathway instead of someone quietly signing up for a third-party tool. ZTNA and the wider cybersecurity controls keep unauthorised AI use out. The CRM agent replaced a manual meeting-prep process that used to take the sales team hours each week.

Pathfinder Asset Management · 2023 → 2026

Enterprise CRM for Pathfinder Asset Management

<90: Days to delivery
30%: Under budget
15,000+: Clients served

Problem: Pathfinder Asset Management, part of the wider group, needed a full enterprise CRM to serve 15,000+ clients. The timeline was tight.
Approach: Kicked off and managed the delivery of a Dynamics 365 CRM build for Pathfinder. Scoped it against their actual workflows rather than a generic template. Ran it fast and controlled the budget closely.
Outcome: Delivered in under 90 days, 30% under budget, serving 15,000+ clients. Also used the engagement to architect the current-state and future-state CRM structure for the wider group, covering multiple legal entities and customer segments.

Helm Wealth · 2023 → 2026

Cybersecurity Programme

Problem: No formal cybersecurity baseline. The firm was growing through acquisitions, adding attack surface with every integration. Risk and compliance needed assurance, and the board needed reporting they could act on.
Approach: Rolled out a full enterprise cybersecurity programme from scratch. SIEM for event monitoring. Secure web gateway. Endpoint detection and response. Phishing training across the firm. Firewalls, identity management, MAM via Microsoft Intune, and disaster recovery processes. Aligned the whole thing to the CIS controls baseline and ran tabletop exercises to test it.
Outcome: The firm went from no formal cybersecurity posture to a CIS-aligned programme with continuous monitoring, trained staff, and a disaster recovery plan the board has seen tested. Incident volume dropped. Platform availability went up. Risk committee reporting now covers cybersecurity as a standing item.

Helm Wealth · 2023 → 2026

Data Platform and Reporting Automation

Problem: Every team was doing reporting by hand in Excel. No single source of truth. Finance, compliance, operations, and investments were all maintaining their own spreadsheets with their own numbers.
Approach: Built a new data platform on Snowflake with Coalesce for data transformation and Power BI for reporting. Designed governed datasets and standardised models so the numbers matched across teams. Replaced the manual Excel work with automated dashboards and reporting pipelines.
Outcome: Frontline workload dropped by 25% across finance, compliance, operations, and investments. Reporting cycles shortened. The data platform also became the foundation for the AI programme: you can't run agents against data you don't trust, and now the data is governed.

Helm Wealth · 2023 → 2026

Five Acquisition Integrations

Problem: Helm grew through acquisition. Five separate businesses (Newton Ross, Salt Funds Management, CP Wealth, Octagon Asset Management, and Summer KiwiSaver Fund) needed to be folded into the group technology platform. Each one brought its own systems, identities, data, vendors, and service arrangements.
Approach: Led the technology workstream for all five, from pre-acquisition due diligence through to post-acquisition integration. Each integration covered identity and access management, M365 migration, CRM consolidation, data migration, and vendor rationalisation. Ran cutover planning with rollback plans and rehearsals for every one.
Outcome: All five integrated with no material disruption to operations. The group now runs on a single platform with consistent identity, security, and service management. Vendor contracts consolidated, SLA adherence improved, and run cost reduced.

Helm Wealth · 2023 → 2026

Vendor Reset and Service Resilience

Problem: Incident resolution was averaging five days. The managed service provider wasn't meeting SLAs. The technology team was spending more time chasing tickets than doing anything productive.
Approach: Reset the vendor arrangements. Replaced the underperforming MSP. Negotiated new SLAs with clear accountability. Introduced change discipline, problem management, and proactive capacity reviews.
Outcome: Incident resolution dropped from five days to about 2.5 hours. Incident volume fell. Platform availability improved. The technology team got their time back.

Helm Wealth · 2023 → 2026

Technology Strategy and Board Investment

Problem: Technology investment was minimal. The board had no visibility into technology risk or opportunity, and no confidence that spending more would deliver returns.
Approach: Built the technology strategy and multi-year roadmap from scratch. Wrote the benefits cases. Sequenced delivery to show early wins that built board confidence. Reported regularly to the board and risk committees on technology, cybersecurity, and service stability.
Outcome: The board backed the strategy with a tenfold increase in technology investment. That investment funded the AI programme, the data platform, the cybersecurity programme, the CRM rollouts, and the infrastructure modernisation. The strategy work earned the money. The delivery programme spent it well.

Be Intelligent (now part of PwC) · 2022 → 2023

Managed Services from Zero

Problem: A Salesforce consultancy with no managed services practice. Revenue was entirely tied to one-off implementations. The market was crowded and the firm had no recurring income.
Approach: Designed the managed services practice from scratch: the delivery model, the pricing structure, and the SLAs. Took it to market. Turned one-off Salesforce implementations into recurring engagements with success plans and renewals. Built the QA gates and support workflows that kept client satisfaction high. Ran both the delivery and the commercials.
Outcome: Revenue grew 350% in five months. Client satisfaction hit 4.8 out of 5. The practice grew to generate 25% of Be Intelligent's total revenue — a recurring-revenue engine in a market where most competitors were still selling project-to-project.

Fresh Direct Group · 2021 → 2022

ERP Rescope and Platform Modernisation

Problem: A $200M food distribution group running more than ten operating businesses on ageing platforms. The ERP was overscoped and overpriced. Warehouse and dispatch workflows were slow. Management reporting was done by hand. The board had limited visibility into technology risk.
Approach: Migrated 250+ users from on-premise infrastructure to Azure. Rescoped and renegotiated the ERP, aligning the solution to what the business actually needed rather than what the vendor wanted to sell. Redesigned the warehouse and dispatch workflows. Moved management reporting onto governed Power BI models. Reported to the board on strategy, risk, and investment priorities.
Outcome: $450K saved on the ERP (from $1.4M down to $950K). National dispatch times shortened by more than 20 hours a day. 250+ users on cloud infrastructure. The board had technology reporting for the first time.

Datacom · 2018 → 2020

Dynamics 365 Implementations and the HR Practice

Problem: Enterprise clients replacing legacy ERPs they had outgrown. Weak controls, manual processes, and no modern HR platform. Datacom also had no Dynamics 365 HR practice.
Approach: Led Dynamics 365 Finance and Operations implementations end to end, sitting inside client finance, operations, and HR teams from scoping through go-live. Built Datacom's Dynamics 365 HR practice in partnership with Microsoft, including a custom DataPay integration. Trained client teams to own their configuration and reporting after go-live so they didn't need us permanently.
Outcome: $2M in qualified HR pipeline generated with Microsoft. Multiple successful ERP go-lives with low defect counts. Client teams that could run their own systems after handover.

Projects

Apex Predictor

Active

A bot that runs a Fantasy F1 team end to end. Under the hood: an integer linear programming solver that searches 1.4 million possible teams in under two seconds, a price simulator that stops it thinking short-term, and a language model that reviews every pick before it locks in. Started as a single Claude Code prompt, then refined the hard way.

DRONE: BREACH

Active

A browser-based hacking game where an AI called Mother is the antagonist. She watches what you do, adapts her defence in real time, and escalates when you get close. An experiment in adversarial agent design: how far can you push autonomous decision-making before the model starts making bad calls? Built with Claude, playable in a browser.

Pit Lane to Production

Active

A Fantasy F1 league on Substack where frontier AI models compete as team principals. Claude versus GPT versus Gemini versus the rest. Each round, the models analyse live data, make their picks, and explain their reasoning. The point is what their decision-making reveals about how these models handle uncertainty and trade-offs under pressure.

Two-tier model routing

Shipped

A routing pattern that sends simple queries to a cheaper model and complex queries to a frontier model. When you're running agents in production, the economics matter. Not every query needs the most expensive model.

Work with me

Let's have coffee. Book a discovery call.

Get in touch