Home › Our work › AI Business Intelligence
Our work / AI Business Intelligence
Making sense of your sector’s news, without reading all of it
A precision manufacturer sits at the centre of a commercial network it cannot see. We built a system that reads that whole network every day and answers questions about it in plain English, with a citation behind every claim.
The problem we were asked to solve
Ask a manufacturing company who their customers are and you get a clean answer. Ask who those customers’ competitors are, which of their vendors was just acquired, which client announced a plant expansion last month, or which two names on the approved-supplier list quietly became the same company, and the answer costs a week of somebody’s time.
The information is out there. It sits in federal registries, in trade press, in filings, in the companies’ own announcements. What a company lacks is the hours to read it all, and a single place where a question can be asked of it.
That was the brief: make the latest news in a client’s sector understandable, without asking anybody to read the news.
What we built
The system has two halves. Keeping them separate is what makes it trustworthy.
The first half is a knowledge graph. Every company in the client’s world becomes a node, seeded from federal registry data, which gave us 10,114 companies with their real identifiers, addresses and corporate parents. On top of that we draw the client’s own curated lists: 473 vendors, 187 customers and their competitors, connected by 8,734 verified links. Those edges are hand-checked, and they are the gold standard everything else is measured against.
The second half is retrieval. A scheduler works through the vendor and customer lists on a rotation, pulls sector news for each name, and files every article as a source document attached to the companies it mentions. When you ask the system a question, it finds the passages that actually bear on it and answers from those, quoting them back to you.
Why you can trust what it tells you
This is what clients ask about first, and rightly. Three rules do most of the work.
A relationship needs a verbatim quote. When the system reads an article and concludes that company A supplies company B, it has to produce the exact sentence from that article that says so. The edge is drawn only when that quote exists. That one rule accounts for most of what people have in mind when they worry about an AI system inventing things.
Names resolve against your list first. A company mentioned in an article is matched against the client’s own roster in order: known alias, then exact match, then close string match, and finally a constrained model call that may pick among real candidates or decline. A name resolved once is remembered, so the next article that spells it differently costs nothing.
Structured facts come straight from the graph. A company’s CAGE code, address, corporate parent and relationship to the client are read directly out of stored properties. The language model writes prose about those facts. The graph is what knows them.
A fourth rule is really a design stance: the system is allowed to say it does not know. Derived conclusions, such as a likely second-tier supply relationship or a probable competitor inferred from a shared customer, are drawn as dashed lines and labelled with the rule that produced them. They are offered as leads to check.
What a client actually does with it
They open a graph of their commercial network and click a company. They get a profile that leads with the relationship to them: customer or vendor, since when, current status. Then recent developments, written as short cards giving the topic, a quote from the source, and why it matters. Then identity and contact details, then a comparison against their own position.
Underneath sits a question box. What happened with this supplier’s new programme? gets answered from the articles the system has actually read, with the links attached. Reports save out as documents. Companies, news and extracted text export as CSV over a date range, so the intelligence can be pushed into a CRM and turned into a call list.
The valuable output is the ten minutes on Monday morning when somebody finds out a client’s parent company changed, in time for the sales call.
Where it runs
Deployed on a server we provision and maintain. The graph database, the application and the language model all run in containers, so the whole system moves as a unit and a new environment stands up in a single deployment.
Cloud infrastructure and the deployment pipeline are part of what we deliver. We size the instance, set up the reverse proxy and certificates, put the application behind authentication, and keep the update path simple enough that a new model or a schema change ships without downtime.
Where a client needs the data to stay inside their own network, the same containers run on their own hardware on a single workstation GPU. For a defence-adjacent manufacturer that option is often the condition of the project existing at all, and the architecture supports both from the same codebase.