The job titles all matched. The intent was in the comments.

Finding names was never the problem. Knowing who was in-market was

Contact data is a commodity. Several vendors will sell an accurate list of every Director of Operations in a region. What none of them sell is the knowledge that one of those directors complained about their tooling under a post yesterday.

Reps were moving between disconnected tools: one to spot intent signals, another to build lists by industry and size, a third for local businesses. Each had its own login, its own export and no shared scoring logic, so a rep's judgement was the only ranking mechanism. That judgement does not survive being handed to someone else.

Extracting the signal reliably fails in specific ways, most of which only appear once a run covers more than a handful of keywords:

• Intent lives in the comments, not the post. The post is the bait; the buying signal is three replies down, in someone else's words.
• Buyers and sellers look identical. Under a post asking for a vendor recommendation, a good share of the commenters are competitors pitching. Both use the same vocabulary, often in the same thread.
• The same person surfaces many times. Without deduplication at the profile level, one prospect becomes six rows and pushes genuinely new names down the list.
• Posts arrive truncated. Feed responses clip the body at "…see more", so the context the model needs is missing exactly when it matters.
• Free-tier rate limits bind everything. Called naively and in sequence, a single run took longer than doing the research by hand.

A screen is cheap. Qualification is not. Do them in that order

The obvious design is a single call: give the model a comment, a profile and a company, ask for a verdict. It does not survive volume. A keyword sweep returns hundreds of comments, and enriching every commenter before scoring spends the profile-lookup quota on people who were never candidates, so the run either stalls on a rate limit or costs more than the leads are worth.

So screening happens on the cheapest signal available, and enrichment is earned. Round one is a deliberately cheap pass over comment text alone: buyer or seller, intent high, mid or hidden, and a seniority check. Survivors, and only survivors, get profile and company enrichment. Round two re-reads each one with that full context and returns a qualification verdict with a confidence score of 1 to 10.

The ordering is the architecture, not an optimisation. Fighting it is the most common reason this class of tool works in a demo and stalls in production.

Two modules that judge, three that cover, one export schema

The platform splits along a clear line. Two modules do judgement work: intent scoring reads the comments under keyword-matched posts, and a Lookup dossier assembles enough context to open a conversation. Three do coverage work, finding companies, contacts and local businesses that match a filter. Everything converges on one scoring step and one export schema.

Lookup runs two pipelines. A person summary chains three collection passes (basic profile and activity feed, then a deep pass for experience, skills and email, then a third that recovers the post content the feed truncated) and merges them into one record the model turns into interests, expertise, a career story and an outreach hook. Company intelligence resolves the domain first, then fires ten sources at once across encyclopaedia entries, news and careers searches, and site crawls, with a registry lookup spanning 15+ countries behind it. The output is a briefing, not a list of links.

The reach modules are held to the same standard: they drop records rather than pad them. A company with no contact after three widening attempts is skipped, and a business with neither an email nor a phone never enters the export.

Full autonomy was never the target. The system decides who is worth a message and hands over the context to write it. A person still decides what it says.