Selected work
Two systems, in detail. Client identities are withheld under confidentiality – the architecture and the reasoning are not.
Five agents,
one content engine.
The company wanted to grow in international markets and assumed it needed more content. We started with their search data instead.
That number reframed the entire engagement. The demand already existed; it was several orders of magnitude larger than the traffic. The pages simply weren’t ranking high enough to be clicked, and weren’t locally relevant enough to deserve it when they were. Producing more of the same content would have added volume to a problem that wasn’t about volume.
The real constraint was that genuinely local, genuinely useful pages – real statistics, real regulations, real competitive context for each market – could not be produced by hand at the scale the opportunity justified.
Live web search per market. Statistics, regulations, market signals.
confidence score
capture date
Discovers competitors per market and extracts their capabilities.
live gap analysis
Clusters real demand into buyer intents, mapped to product and page type.
Assembles the intelligence into a fully on-brand page, structured for the specific intent rather than from one generic template.
Scores output against brand, accuracy and structural criteria before a human ever sees it.
A five-agent system, each stage with a defined output and a quality gate.
Country Intelligence – researches a target market using live web search, extracting statistics, regulations, market signals and competitive context. Every field carries a source URL, a confidence score and a capture date. Nothing enters the system unattributed.
Competitor Intelligence – discovers and analyses competitors per market, extracts their capabilities, and diffs them against the client’s feature library to produce a live gap analysis.
Keyword & Demand Discovery – clusters real search demand into distinct buyer intents, then maps each cluster to a product and page type. Crucially, this prunes: intents without genuine demand don’t become pages. Publishing pages nobody searches for is how programmatic content becomes a liability.
Page Generation – assembles the intelligence into a fully on-brand page matching the existing site exactly, structured for the specific intent rather than from a single generic template.
Quality Assurance – scores output against brand, accuracy and structural criteria before a human ever sees it.
Every generated page is created as a draft on the live production site. Nothing publishes automatically. A person reviews, edits and approves before anything is visible to a customer.
The intent map itself is a second, earlier gate – the plan for what pages will exist is approved by a human before generation begins. This matters more than the output gate: it’s cheaper to reject a plan than to review fifty pages that shouldn’t have been made.
Re-runs update pages in place rather than creating duplicates, so the system can be re-run safely as source data changes.
Page production went from days of manual research and writing to minutes. Because the shared template is central, an improvement to it upgrades every page already published.
The engagement also established something the company didn’t previously have: a measurement framework. Per-market baselines for position, impressions, clicks, click-through rate and demo requests, tracked against the pages as they go live.
On outcomes, we’ll be precise. Search rankings move on Google’s timeline, not ours – indexing takes days to weeks, and meaningful ranking movement takes roughly ninety days. At the time of writing, the system is delivering pages and the measurement framework is live; ranking outcomes are still maturing.
We could describe projected multiples here. We’d rather describe what’s true.
The interesting decision wasn’t technical. It was refusing to treat page volume as the goal.
The theoretical page count for this content matrix ran into the tens of thousands. Building all of it would have been straightforward and would have actively damaged the client – thin, low-demand pages are penalised, and at scale they can drag down the pages that matter. The demand-pruning rule was the most valuable thing in the architecture, and it exists to make the system produce less.
Most content-generation systems optimise for output. The constraint is what makes this one safe to run.
Lead intelligence,
without the seat licences.
The sales team was researching prospects by hand. Find a company, open its website, work out whether it plausibly fits the product, hunt for someone with a relevant title, then write an opening message from scratch. Repeat.
It’s the kind of work that looks like a data problem and is actually a judgement problem. The slow part isn’t finding companies – plenty of databases do that. The slow part is deciding whether a given company is worth anyone’s time, and that decision was being remade from scratch, inconsistently, by different people.
A lead intelligence platform that automates the judgement, not just the lookup.
Discovery – three ways in, because prospecting isn’t one workflow. Open search across the live web and B2B data providers. Spreadsheet import, with automatic column mapping across sixteen fields so a list from any source is usable without reformatting. And a lookalike engine: point it at existing clients and it builds targeted queries seeded from their profiles, returning prospects badged with which client they resemble.
Enrichment – every company is researched live before it’s scored, so the assessment reflects what’s true now rather than what a database recorded eighteen months ago.
Scoring – each prospect is rated against a product-specific ideal customer profile, with a written explanation of why it fits. The explanation matters more than the score. A number tells a salesperson to make a call; a reason tells them how to open it.
Outreach – the named decision maker is surfaced with their title, alongside separate drafts for email, LinkedIn and X, each written to that platform’s format rather than the same paragraph pasted three times.
Client libraries are scoped per product, so one product’s ideal customer never contaminates another’s scoring.
Early runs produced company websites that didn’t exist.
The model, asked to summarise a company from search results, would occasionally fill a gap with a plausible-looking URL. Nothing in the output flagged it. To a salesperson, an invented domain is indistinguishable from a real one until they click it – and by then the lead is in the pipeline and trust in the whole tool is damaged.
Two fixes. Every scoring prompt carries explicit constraints against asserting anything not present in the retrieved sources. And every URL is validated at DNS level before it reaches the interface, so a fabricated domain fails a real check rather than a stylistic one.
This is the difference between a demo and a tool people rely on. A demo shows you the good output. A production system has to be honest about what it doesn’t know – and be built so that when it’s wrong, something catches it before a person acts on it.
Prospect research moved from manual work to a reviewed queue. The team’s judgement is now applied once, encoded in the scoring criteria, instead of re-improvised per prospect – and every lead arrives with a stated reason it was surfaced, so a salesperson can disagree with the system rather than simply trust it.
Deployed behind team authentication, running against four external data and AI services, with no API credential ever reaching the browser.
