How we built one AI brain that coaches and closes for Pitho.

The challenge

Pitho is built for high-ticket closers, the people who sell $3K to $25K offers inside the direct messages of Instagram and LinkedIn. The hard part of that job was never typing.

The expertise that made it work was scattered. A proven 14-stage closing framework and a library of real winning conversations lived in one founder's head and a handful of playbooks. The leads lived in a CRM, the conversations somewhere else, and the live inboxes across four different messaging platforms. None of it could help a closer in the exact moment they needed it: knowing which of fourteen stages a conversation was in, what to say next in a voice that did not sound like a bot, and which of a few hundred half-finished threads was actually worth chasing before it went cold. That was the problem.

One brain, where the closer already works

We centralized all of it into one brain and put that brain wherever the closer already works. It pulls the real conversation straight from the platforms through Unipile, a unified messaging layer that turns LinkedIn, Instagram, Gmail, and Outlook into a single API, so Pitho works off real threads instead of scraping anyone's browser.

The intelligence lives in a Python and FastAPI backend that streams its replies back token by token, on top of a single Supabase Postgres database that handles auth, per-user isolation, and realtime. Four moving parts, one core.

The reply is not one model

The reply is the part most people underestimate, because it is not one model writing back. A classifier reads the full thread and the CRM tags and decides which of the 14 framework stages the conversation is in. At the same time, a Pinecone search pulls patterns from two knowledge bases at once, one of playbooks and one of real closed deals, both embedded with OpenAI.

A generator then writes the reply in the founder's voice, locked to that stage. A judge scores it on eight dimensions, things like stage fit, voice, mirroring, and whether it actually sounds human, and if the draft falls short a refiner rewrites it and runs it back through, keeping the best version and discarding any rewrite that scores worse. The whole loop runs on Claude, with the cheap Haiku model doing the fast work and Sonnet handling the heavier reasoning, so cost stays where it belongs.

What it means in practice is simple. The closer opens a thread and gets a reply that already knows where the conversation stands, sounds like the person who built the framework, and arrives with its stage and its quality score attached.

It hunts, it does not just answer

The same brain does more than answer in the moment. It goes hunting. A background scanner pages the closer's entire inbox through Unipile, scores every single contact from cold to qualified, and ranks them into saved lists: the leads worth re-engaging, the people who asked but never booked, the past customers worth a second touch.

Those lists become reactivation campaigns, one message drafted per person in the closer's own voice, reviewed and approved before anything goes out, with sends paced so Instagram and LinkedIn never flag the account. The system stops being a thing you ask and becomes a thing that tells you who you forgot.

One brain, three front doors

Sitting over all of it is one agent that ties the product together, a single tool-using Claude loop that knows each user's scored contacts and what they sell, and can build lead lists, qualify people, draft, send, and update the CRM on its own. That exact same agent runs behind all three ways into Pitho, so the Chrome extension, the web inbox, and the Telegram bot are one brain rather than three separate apps.

The extension reads whatever DM tab the closer is on and answers in a sidepanel. The inbox runs live on Postgres realtime. The Telegram bot takes text, voice notes it transcribes with Whisper, and photos, so a closer can run the entire operation from their phone, walking down the street, by saying find me the warm leads I never followed up with, draft the reactivations, send them.

Safe by design, not by trust

Because that brain acts on real accounts and sends real messages, we made it safe by design rather than by trust. Every member's data is walled off at the database layer through row-level security, so no one's leads, conversations, or scores can ever surface in anyone else's session, and the web app holds no keys, it simply proxies to the backend.

Nothing irreversible ever fires on its own. When the agent wants to send a DM or change a lead, it stops and asks, and the action only goes out when the closer taps confirm, in the browser or on a Telegram button, and only from the account that asked for it. The closer moves fast, but nothing leaves their account without their name on it and their finger on the button.

One product, in production, owned

It all runs in production: the messaging, the retrieval, the four-agent pipeline, and the proactive scanner, sitting on one database behind per-user isolation, with every model call routed through a single backend the company controls. One platform that coaches the closer in the moment, hunts down the leads they let slip, and answers from a phone or a browser with the same brain behind it.

This is what owning the system looks like, the orchestration layer that a stack of rented, per-seat tools could never be. Pitho owns every line of it, deployed on their own infrastructure, and it keeps getting sharper as the conversations pile up.