Lucas Peruzzo

A WhatsApp support platform for the real-estate market. It talks to the lead, works out what they want, answers questions from the property's own documents, books the visit and hands a qualified lead to the right broker — on its own, from the first hello to a time on the calendar.

• n8n
• Supabase
• Convex
• Cohere Rerank
• WhatsApp Cloud API
• WAHA

• It waits for the person to finish typing

  Nobody writes everything in one tidy message. Incoming messages land in a queue and a 20-second timer restarts on each new one, so the agent replies once — after the person has actually stopped — instead of firing four half-answers at four fragments.

  Supabase queue · 20s debounce reset per message · batched into one turn

• Specialist agents, not one giant prompt

  Every stage of the conversation has its own agent with a single job, instead of one prompt trying to cover everything. Narrower scope means sharper answers — and it holds up on cheaper models, which is what keeps the cost per conversation survivable at volume.

  multi-flow orchestration in n8n · data shaped between every stage

• Answers come from the client's own documents

  The client's files are indexed automatically. When someone asks something specific, the system retrieves the right passages, reorders them by real relevance and answers from the text — so it says what the document says instead of inventing a floor plan.

  RAG · automatic Google Drive indexing · vector search + Cohere Rerank

• Hears audio, reads images, sends media

  The lead can send a voice note or a photo and the system understands it. It answers in whatever form makes sense back: text, audio, an image, or the development's own brochure.

  audio transcription · image understanding · media and document delivery

• Books the visit straight into the calendar

  The agent checks the broker's genuinely free slots, offers real options and closes the appointment inside the conversation — no 'someone will get back to you'.

  reads and writes availability · confirmation inside the flow

• Hands the lead over in the right order

  A two-tier rotation: managers first, then brokers, first in first out. If nobody is on duty the lead is held rather than burned, and released every five minutes as people come back — so one broker never collects three leads while a colleague gets none.

  two-tier FIFO rotation · hold with release every 5 min

• Moves the deal through the pipeline by itself

  The agent changes the Kanban stage as the conversation develops. The funnel organises itself up to the handover, which means the board reflects what actually happened rather than whoever remembered to drag a card.

  stage changes triggered by the agent · business rules inside the flow

• Brings back the ones who went quiet

  A stalled conversation is not a dead one. Follow-ups fire on a schedule tied to the funnel stage and to how long the lead has been silent.

  follow-ups scheduled per stage and per inactivity window

• Two WhatsApp channels

  Runs on Meta's official Cloud API or on an unofficial connection, whichever fits the client — the same agent sits behind both.

  WAHA · WhatsApp Cloud API (Meta) · webhooks

• An operations panel that stays fast

  Volume, funnel stages and performance, still loading quickly once the data has piled up — because the numbers are aggregated ahead of time instead of counted on every page view.

  pre-computed aggregations in the database

• The whole application, not just the agent

  The database, the login, the container and the domain are mine too. From the first data model to the system answering in production.

  Convex · Clerk · Docker · Coolify · Vercel · DNS and domain

A platform for recruiting commercial partners. The agent prospects brokers and agencies to sell as partners and, the moment the partnership closes, it changes role: it stops selling and becomes a product consultant, with every document on hand to answer anything on the spot.

• n8n
• Convex
• Clerk
• WhatsApp Cloud API
• Docker
• Coolify

• The agent changes job halfway through

  One funnel stage flips the operating mode: the person being prospected becomes a partner, and the same channel starts delivering product information instead of a sales argument.

  agent context swap triggered by funnel stage

• A product consultant that reads the documents

  Once they are a partner, the broker can ask anything about the product and get the right answer — pulled straight from the official material, with the document attached when that helps.

  RAG over indexed Drive files · document delivery orchestration

• The same engineering, far simpler to run

  The same multi-agent architecture underneath, behind a stripped-down interface. The difference here is not what it can do — it is how little it takes to set up.

  multi-agent setup reduced to interface-level configuration

• Audio, images and the same smart queue

  The conversation core is shared with Imobium: message batching, audio and image understanding, follow-ups and a filtering pass before anything is sent.

  debounced queue · transcription · vision · automatic follow-ups

• Under review for official Meta partnership

  Once approved, the client connects their own WhatsApp inside the platform in a few clicks — with no technical step in the middle.

  embedded onboarding via WhatsApp Cloud API

• Built and shipped by me

  Same infrastructure base: database and functions on Convex, login and permissions on Clerk, the application packaged in a container and published on its own domain.

  Convex · Clerk · Docker · Coolify · Vercel · DNS and domain

• A deliberately narrow scope

  There is no calendar and no lead rotation, and that is a choice: the problem here is recruiting and informing a partner, not booking a property viewing.

  designed for a partnership funnel, not a direct-sales one