The public experience brings travel services and enquiry entry points together, letting visitors browse packages or start a conversation with a consultant.
Separate recommendation, email drafting, and feedback-routing agents use structured output parsers. A Gmail approval step either releases the final email or returns feedback to the draft agent.
A scheduled branch generates text and an image, stores the image in Drive, and writes a draft to Sheets. A separate approval trigger sends the approved image and text to Telegram and records publishing or error status.
FastAPI BackgroundTasks process PDFs into document-scoped embeddings. The agent can retrieve context, look up weather, evaluate allowlisted arithmetic, and extract skills using a predefined vocabulary.
FastAPI
LangChain
Gemini
pgvector
This is a single tool-using agent with FastAPI background tasks, not a Celery queue or a general multi-agent system.
An incoming webhook reaches a Gemini-powered agent with a Pinecone vector-store tool and Hugging Face embeddings. Its output is returned through the webhook response node.
The connected execution path is WhatsApp trigger → AI agent → WhatsApp send. The agent also has conversation memory and a Gmail tool; extra disconnected nodes are not presented as active features.
A FastAPI chat endpoint discovers tool schemas through a FastMCP client, asks Gemini for a routing decision, and executes the chosen tool through MCP. Notes and trainer prompts are exposed separately as resources and prompts.
A ColumnTransformer one-hot encodes location and passes numeric fields into LinearRegression. A held-out split is used for MAE and R²; joblib saves the fitted preprocessing and estimator together.
Python
Pandas
scikit-learn
joblib
No reproducible benchmark values or real-world valuation reliability are claimed.
