Projects

AI triage and knowledge enrichment for ITSM, built to be auditable from the first line of code. Every classification is logged and explainable, because automation that can't show its working doesn't survive contact with ITIL. Production triage accuracy currently runs at 99.8% on live customer traffic.

Related writing: The Moat Was Never the Code

Legacy estates moved into Azure properly, not lifted and shifted. Containerisation, CI/CD and Infrastructure as Code replaced environments nobody dared touch with ones that rebuild from source. Lift-and-shift is faster on paper; you just pay for it every month afterwards.

Legacy ITSM documentation rewritten into summaries, FAQs, how-tos and troubleshooting guides, then published to a modern customer portal. Search runs hybrid vector and full-text retrieval, reranked by Cohere, so users get the most relevant answer rather than the closest keyword match. The hard part isn't generation; it's making twenty years of documentation findable.

A transformation studio that turns monolithic Word and PDF documentation, screenshots included, into structured Quick Starts, FAQs, How-Tos, Cheat Sheets and Troubleshooting guides. Built on a hard rule: the AI restructures, it never invents. Automated validation catches unverifiable URLs, paths and settings before anything ships, because commercial documentation can't contain fiction.

Related writing: I Run Three Sprints a Week. I'm Not Sure They're Sprints.

Traditional contact centres rebuilt as AI-augmented service hubs. Voice AI handles call classification and intelligent routing; humans keep the judgement calls. The human-in-the-loop design isn't a compliance fig leaf: agents get better data, callers get faster answers, and nobody talks to a robot pretending otherwise.

Mobile-first tools for field engineers who fix boilers, not interfaces. Real-time job status and guided workflows cut cognitive load, while compliance and reporting get captured as a by-product of doing the job rather than a chore bolted on at the end. If an engineer needs training to use it, it's failed.

ETL pipelines that turn raw operational logs into something AI can actually use. Unifying messy ITSM datasets is the unglamorous prerequisite for everything else on this page; without reliable data engineering, "applied AI" is just a demo. This is the plumbing that makes the rest work.

Related writing: The Moat Was Never the Code

Real-time reporting frameworks giving service leaders a live view of workloads, service levels and activity trends. Built on a simple principle: if a manager has to ask for a report, the reporting's already failed. Decisions get made in the moment, from data that's seconds old rather than days.

Structured handling for complex customer issues, unifying operational data, evidence and compliance requirements into auditable workflows. Consistency is the point: two customers with the same problem should get the same outcome, and the audit trail should prove it.

An AI-assisted repository for tender responses, governance documentation, ISO certification evidence and compliance artefacts. Scattered compliance knowledge becomes a searchable, reusable asset, so tender responses start from the best previous answer rather than a blank page, and certification evidence stays current instead of being rediscovered in a panic each audit cycle.

A campaign management application integrating directly with the Google Ads and Microsoft Advertising APIs. Campaign builds, budget control and performance reporting across both networks from one interface, replacing manual platform-hopping with auditable programmatic control. The platforms' own UIs are built to spend your money; an API client is built to spend it well.

Web platform for a mobile mechanic business in the Black Country. Mobile-first booking flows, vehicle lookup and AI-assisted enquiry capture, tuned hard for local search. Small business sites don't need clever; they need to rank, load fast and convert a stressed driver into a booked job.

AI-assisted QA across field service job records, surfacing risks and inconsistencies for human review, including repeat-visit analysis and chargeable versus non-chargeable outcomes. In regulatory-sensitive environments the AI flags and people decide. Transparency isn't a feature here; it's the licence to operate.

Related writing: I Run Three Sprints a Week. I'm Not Sure They're Sprints.

Retrieval-augmented generation built for production rather than demos. Incident, problem and knowledge records flow through a hybrid vector and full-text pipeline with HyDE query expansion and Cohere reranking, so retrieval holds up on the imprecise, conversational queries real users actually type. Most RAG quality problems are retrieval problems; this is where the gap between demo and production gets closed.

Related writing: Beyond Basic RAG: HyDE and Cohere Reranking

E-commerce build for a dog supplement brand: three products, Stripe checkout, Sanity-managed content. The brief was deliberately minimal, because a three-product store doesn't need a platform. It needs to load fast and take payment without friction.

Technical SEO and generative engine optimisation across multiple production sites. Structured data unified into a single @graph with entity disambiguation, per-route metadata, server-side rendering for crawlability, and auto-generated llms.txt files that stay in sync with the database via build hooks. Classic SEO gets you found by Google; GEO gets you cited by ChatGPT, Claude and Perplexity, and most sites optimise for one while ignoring the other. Lighthouse SEO score on the most recent remediation: 100.

AI analysis of customer feedback in ITSM contexts, tuned to catch nuance and outliers rather than averaging star ratings. A thousand "fine" responses tell you less than one furious outlier, and the system's job is to find that one. Privacy-aware and explainable by design.

Conversational self-service for IT support, focused on capturing better requests rather than deflecting users. Intent understanding plus structured handover means a human agent starts with context instead of a blank ticket. Chatbots fail when they're built to avoid customers; this one's built to brief colleagues.