nearby
An autonomous pipeline from territory to booked meeting: scan an area, diagnose every business in it, score the fit, run the outreach — and put a human in the room for the close.
Pipeline generation forces a bad choice: human outreach that does not scale, or automated outreach with no judgment — the spray everyone deletes. nearby asks where judgment is actually required, and answers: in exactly five places. Everywhere else, deterministic code is faster, cheaper, and correct every time.
A territory is swept via the places API: every business, its profile, its reviews — landing as rows in Postgres, categorized deterministically against a taxonomy.
An auditor agent reads each business's reviews and presence the way a consultant would — finding operational pain points and estimating what they cost per month. One audit per business, run once, reused everywhere.
A lightweight model scores each audit against the defined offers and proposes the angle; SQL — not a model — selects who clears the threshold. Judgment where judgment is needed, arithmetic where it is not.
A drafter writes outreach grounded in the diagnosis — the specific pain, the specific number. Inbound replies are classified by intent with confidence scores; only interested replies and objections get a drafted response.
The pipeline ends in a calendar booking with a person. The machine does the hours; a human does the close — a design decision, not a limitation.
Stack: plain Python, Postgres, and cron. Every agent is a stateless function — read rows, call one API, write rows. Agents never call each other; they communicate through table state. No workflow framework, deliberately.
The fashionable design — agents delegating to agents — was considered and refused. Filtering by geography is not a judgment call; it is a SQL WHERE clause. Models as middle managers trade fast, free, exact operations for slow, expensive, unreliable ones.
The first design ran a full model audit for every business-and-offer pair: at 5,000 businesses and four offers, 20,000 expensive model calls, roughly $420 per sweep. The redesign audits each business once, then scores each offer against that audit with a lightweight model: about $105 for the same territory, around 2¢ per business. Same judgment, a quarter of the bill.
The first stage scans and audits an entire territory without sending anything; the operator reads the diagnosed market and picks the niche before a single email leaves the system. Only then does outreach switch on, one stage at a time.
nearby's first deployment is our own pipeline — we are its first operator, and its territory diagnostics feed how Alot Labs itself finds the businesses we can genuinely help.
The same architecture pointed at your market and your offer: your definition of a good-fit customer, your outreach voice, your calendar at the end of it — running in your infrastructure with every stage inspectable in the database.