The problem: at a trade fair a consultant gets about 90 seconds with a decision-maker. Everyone pitches, nobody remembers the pitch, and the follow-up email is generic.
What we built: - /capture (phone): snap a badge or type a company. Hallway Hunter finds the right legal company in the Finnish trade register (PRH), checks public tenders (TED) and the company's own website, and returns a 120-word brief with numbered sources: what they do, recent signals, three openers, what not to say and the likely objection. Measured on live cold runs: 5 to 9 seconds from badge to brief. Names are never stored, only the role. - Save what they actually said (a voice note) and, optionally, how many teams work like that. - Overnight, Velho's AI analysts draft one page for that company: their own words back to them, one costed process (an editable formula labelled as an estimate, scaled by the teams they named), what we would check first, and one short yes/no question signed by a named sender. - /hq: a person reviews each draft with its compliance checks and a reader test (an AI reading the page as the company's COO), then approves. Approving only writes a draft email to our own inbox. Nothing is ever sent automatically.
How it works (and what's AI about it): Next.js 16 and SQLite. Claude writes the scout brief, the overnight research, the page script, and runs as a skeptical reader. Hours and euros are computed in code, not by the model. A citation sanitizer drops any claim without a source, subsidiary or namesake tenders are kept out of the signals, and a compliance check holds any page that states an uncited conclusion as fact.
Built for Matrix OS: the night path needs a machine that stays on, so Hallway Hunter is packaged as a Matrix app (matrix/matrix.json). On the box a loop runs the night path at 22:00, and at 07:30 Matrix delivers a digest of overnight pages waiting for approval to our own Telegram chat (never to a company). The night path starts in the background and the page polls, because Matrix's gateway cuts any request after 30 s. A shared Agent Ops board, a Matrix app reading Matrix Postgres tables, shows Hallway Hunter's and Ringo's runs and approvals live. matrix/deploy.sh packs, uploads, builds and registers everything (a dry run unless --apply). It is deployed on our Matrix box: the app is in the launcher with the demo drafts from the video waiting in /hq, and the night loop is running there.
Honest limits: the reader test still scores "would reply" at 3 to 5 out of 10, because one team's saving looks small to a COO; the scaled figure only appears when the person in the hallway says how many teams work that way. The demo quotes are invented and labelled as such. On stage we ask a judge for one real annoyance at their company and record it as the hallway note.
What's next: a real reply button to the consultant's inbox, more public sources per company (news, annual reports), and measuring real replies instead of a simulated reader.
