Trato

Inspiration

Freelancers lose money in the gaps between a conversation and a payment: scope creep that was never written down, change requests agreed by chat, deliverables that the client "never approved", refunds argued from memory. We wanted a workspace where every phase of a project leaves documentary evidence, and where getting paid is tied to that evidence.

What it does

Trato is a workspace for freelancers that turns project documents into proof of each phase:

• Project start: upload the initial conversation or agreement. Trato reads it and proposes deliverables and staged payments, using published reputation of the client as context.
• Change requests: when the client asks for something new, Trato compares the request with what was agreed, drafts the scope, conditions and effort (with quotes from your own documents), and you confirm your rate to calculate the additional fee. The addendum is generated as a versioned PDF and sent from Gmail.
• Client answers: replies are read from Gmail; Trato analyses whether the client accepted or asked for adjustments, and each new version keeps its own PDF, sending and acceptance.
• Payments with PayPal: each deliverable or accepted change produces a PayPal checkout. A change or stage is only marked as paid when the backend verifies the PayPal capture — never because the user returned from PayPal.
• History and evidence: one timeline per project (start, initial sources, requests, addenda, acceptances, payments) and a project map, so any phase can be audited later.
• Rehearsals: practice a difficult client conversation with an AI client and get feedback grounded in negotiation guides.

How we built it

• Frontend: vanilla JS modules, bilingual UI (ES/EN), Lucide icons.
• Backend: Python FastAPI (auth, business rules, Gmail, PayPal), Supabase (Postgres with row-level security, Auth, Storage).
• AI: Google Gemini through LangChain/LangGraph workers on a Redis queue; every agent returns JSON validated in code, and citations are checked literally against the original documents.
• Documents: a file processor extracts and indexes PDFs into Weaviate (vector search per project, isolated tenants).
• PayPal Sandbox: Orders API with server-side verification and refund webhooks.
• Google Cloud: Cloud Run (web), Compute Engine VM (API, workers, Redis), Artifact Registry.

Challenges

• Making AI useful without inventing facts: strict JSON contracts, evidence checks and visible errors instead of silent fallbacks.
• Payment correctness: idempotent checkouts, payer verification and immutable paid changes.
• Keeping each version of an addendum, its PDF and its acceptance auditable.

What we learned

Grounding the AI in the user's own documents, and treating payments as verified facts, is what makes a freelancer trust the tool.

Try it

Demo: https://trato-web-68896425401.us-central1.run.app Example case for judges (English): ejemplos/cafe-aurora-en in the repository.

Built With

• cloud-run
• compute-engine
• docker
• fastapi
• gmail-api
• google-cloud
• google-gemini
• javascript
• langchain
• langgraph
• paypal
• paypal-sandbox
• postgresql
• python
• redis
• supabase
• weaviate