AI and data science
Over a decade building and leading data science and AI work, including generative AI systems running at a scale most teams have not had to deal with yet. Brought to bear on a specific problem rather than a general engagement. Most of what I am asked for falls into three shapes.
AI strategy and roadmap
Where AI actually pays off in your business, and where it will burn budget. Use case identification and ranking, build versus buy, sequencing, and an honest read on what your data and your organisation can currently support.
- Opportunity assessment with a ranked shortlist, not a long list
- Build, buy or partner decisions with the trade-offs written down
- A roadmap with sequencing and dependencies, sized to your team
Generative AI at production scale
Almost everyone can get an LLM to do something impressive once. The problems that matter start later: when the same pipeline has to classify millions of items a month, across dozens of markets and languages, unattended, with numbers going to clients who will notice a two point move.
That is the work I take on. Not whether to use an LLM, but what it costs to run one properly and how you know it is still right six months in.
- Evaluation that survives volume: gold standard sets, agreement measured against human coders, and detecting quality drift before a customer reports it
- Cost engineering: model tiering and routing, batching, and knowing where a cheaper model is genuinely indistinguishable and where the saving is an illusion
- Model and vendor migration without silently changing what your metrics mean, including backfills and the comparability problems they create
- Behaviour outside English, where output quality degrades quietly and monolingual evaluation will not show it
- The failure modes that only appear in unattended systems: schema drift, partial failures, and the slow ones nobody alerts on
Data science capability
Advisory on building the function rather than the model. Structuring a team, hiring for it, deciding what to centralise, and setting technical direction that survives the first reorganisation.
- Team structure, roles and hiring profiles
- Operating model between data science, engineering and the business
- Technical direction and standards, and how to hold them
Search marketing: keyword research and planning
Separate from the AI work, I take on keyword research and keyword planning for companies running their own Search campaigns. You get a ranked, grouped keyword list with the numbers behind it, not a dashboard to learn.
Most keyword work stops at a raw export: thousands of rows, sorted by volume, with no view on which terms are worth bidding on. The gap is not data, it is judgement about which keywords will actually convert for your business.
I close that gap with a keyword planning tool I built and operate myself, LH Keyword Research. It pulls keyword data from your own Google Ads account, applies machine learning to group and rank it, and I review the result by hand before it reaches you.
How the keyword planning works
Retrieve
Keyword ideas and historical metrics are pulled from your Google Ads account: search volume, seasonality, competition and bid ranges.
Group and rank
Keywords are clustered into related themes, then scored on expected relevance and commercial value using your account's own performance history.
Review and deliver
I go through the ranked list by hand, cut what does not fit, and deliver it as a spreadsheet with a written summary of what to act on first.
What you get
- A ranked keyword list, grouped into themes you can map onto ad groups
- Volume, seasonality, competition and bid range for every term
- A written summary: where the opportunity is, what to drop, what to test first
- A walkthrough call to go over the reasoning
How your account and data are handled
You keep full ownership and control. You grant me access to your Google Ads account, and you can withdraw it at any time from within Google Ads.
The tool reads data only. It never creates, edits, pauses or deletes campaigns, ad groups, ads, keywords, budgets or bids. Any change based on my recommendations is made deliberately, by you or by me, in the Google Ads interface.
It runs locally on my own machine. It is not hosted, there is no login, and I am the only person who operates it. Your data is used only to produce your recommendations: it is not shared with third parties and never informs another client's work. Full detail is in the privacy policy.
Getting started
For AI work, a paragraph on the problem is enough to start. For keyword planning, send the account you want looked at and a sentence on what you sell. Either way I come back on scope, timing and price before anything begins.
Start a conversation