Opportunity radar for a B2B product team

In one paragraph

Every market a B2B company sells into writes its needs down long before they reach a sales call, and in more places than your team can read. This radar reads the market whole. Every morning at 05:00 UTC it goes through 66 live sources – the company’s own customers beside tenders, regulators, competitors, research and the trade press of every industry it serves – and across all 74 it has read 9,114 records, with history back to 2018. On my estimate of ten minutes a record to read it, judge it and log who needs what, that is about 1,500 hours of an analyst’s work, and an ordinary week’s 1,070 new records add some 180 more: four or five analysts doing nothing else. Each signal is weighed against its own source’s history, so the morning surfaces what is new or strengthening, and related signals gather into opportunity cards that point at needs buyers have now and at business models still taking shape. By the start of the working day the cards are on the tray. I built the radar in about five weeks and 319 commits, and since mid-September the product team has worked from that tray.

The problem in business terms

The next line of business for a B2B company is usually in writing before a customer asks for it by name – in lost-deal notes, in tenders where a buyer has committed money, in new rules, in competitors’ job postings, in research grants and in the trade press of the industries it sells into. What stands in the way is volume: over a thousand records in an ordinary week, across dozens of sources and many languages. Read by hand, the market gets read in samples, and samples miss the faint early signals that point at business models still forming. On the day this work started, the people choosing the next bet were choosing from three real cards on their hypothesis board, each written by hand.

Every record, every day, against its own past

Reading a market whole starts as an engineering problem. A feed keeps only its latest few dozen items – one trade outlet produced 50 new records in a single run – so the radar reads daily, before anything scrolls away. A source brings its history when it is switched on, back to 2018 where one exists, because a signal can only be new against a past. And each new kind of source comes in through a gate measured before it opens, so the field widens and the morning queue stays worth your hour.

The second move turns volume into signal. Each record becomes a structured signal – who needs what, quoted and dated – classified as a rate against its own source’s history: NEW, STRENGTHENING, WEAKENING or CONTRADICTING. A source with no history reads UNBASELINED, never NEW, so an unarchived feed cannot pass off its whole backlog as discoveries. A card needs two independent witnesses, and a witness is whoever speaks: three press releases from one press office count once. Cards that rest only on the company’s own customers keep a lane of their own, so the reviewer sees what buyers ask for today beside what the market is starting to ask for by itself.

Three sessions in the engine room

What runs every morning

The run starts at 05:00 UTC, so the queue is ready before the working day begins. Its ten steps go in a deliberate order: the board’s decisions are read first, so the later steps see fresh labels, and the three judges run last, so a judge’s failure leaves the rest of the run intact.

The company’s own customers are a third of what the radar reads; the other two thirds are the market talking among itself.

Every card can be read back to front, and the card page keeps that reading honest. This is the trail of a real card from 8 September, with its subject left out:

The whole archive can also be asked a question. A research assistant the team asks in Slack answers from the same store through a registry of read-only tools – search by words and by meaning, similar cards, a company’s history, competitor moves, tenders, blind spots – at most twelve calls a turn, and computes anything that needs a join inside a sandbox holding one snapshot of the data. It has answered 114 turns so far. Before it reached the team it had to pass a red-team run and a shadow run of 39 approved questions through the real model, with code checking each answer.

The reading bill, in your own numbers

The radar’s limits live in code, where a test can check them

A radar that reads customer text needs its boundaries where a test can hold them. The model inside the run gets no tools: text goes in, schema-constrained structure comes out, and a label outside its vocabulary is dropped in Python, so a hostile page can lie in its content and do nothing more. Customer text is redacted by shape – emails, phones, card numbers, IBANs, tokens – before any model sees it. The research assistant is the one place a model holds tools, and all of them only read, through a database role that sees views and nothing else; an auditor can check those permissions directly.

text in → schema-constrained JSON out
no tools · pinned model version
every call → llm_calls (prompt version, tokens, error)
labels outside the enum → dropped in Python

Deploys are pulled: every five minutes the instance fetches the main branch, builds, checks its health and rolls back to the last good commit if the new one fails. A rollback restores code and never a migration, so each migration has to leave the previous commit working – a line close to a hundred automated security reviews of individual changes have held.

A business-intelligence desk’s work, on a tray every morning

The reading a business-intelligence desk would spend its week on is done before the reviewers arrive, and their hour goes on the cards. In their first week with the queue, the reviewers decided cards faster than the machine produced them – 58 rejections and 27 sends, with two cards left. What changed is the conversation at the gate: a card arrives with its witnesses, its rate against a baseline and the judges’ readings beside it, so the meeting spends its time on whether to bet, and the question of whether the evidence is real arrives already answered. The reviewers also said plainly where the machine fell short. “The first fact, yes,” the Head of Sales said of a card’s evidence, “but the others often seem not quite relevant.” Two plausible causes were measured on production before anything was built, both turned out wrong, and the card now says which of its quotes attest its statement and shows the rest apart. Their verdicts go back into the next morning’s run, so each tray comes closer to what the people at the gate are looking for.

The live constraint has moved to the supply of cards, and that is where the work goes now – more witnesses from outside the company, and judges that earn their way from shadow into the gates on the reviewers’ labels. The purpose stays the same: an ordinary week brings in about 1,070 records, in September about three cards a day came out of them, and the cards worth most arrive before any customer asks – a need the market has started to put money behind, or growing demand for a capability the company once sold only as one-off projects, which is a product asking to be built.

Five weeks. One builder. 66 sources read every morning, some 180 analyst-hours of reading a week done before the working day starts, and 135 opportunity cards put forward – each with its evidence attached.

This case describes architecture and operating discipline. The client, its industry, its product names and every source that would point at either are deliberately left abstract; what carries over to any company is the shape of the radar – read the market whole and every day, weigh each signal against its own past, keep the evidence, and leave the bet to a person.