Case Study — Jarvis

Jarvis — Identity-Aware Personal AI Assistant via WhatsApp

An identity-aware personal AI assistant with real-world execution capabilities

Jarvis is a personal AI assistant designed to manage finances, schedules, and notes through a single conversational interface on WhatsApp. It is built around strong identity awareness, deterministic execution, and extensibility via MCP.

Placeholder: Architecture diagram / system overview image

Context & Problem

Managing personal operations usually requires switching between multiple apps and interfaces. The goal of this project was to explore whether a single conversational interface could safely orchestrate sensitive, real-world actions without becoming a generic or unsafe chatbot.

What Jarvis Does

  • Manage personal financial transactions
  • Register, retrieve, and delete notes
  • List, create, edit, and delete calendar events
  • Check availability and free time slots
  • Maintain conversational context across interactions

All actions are executed deterministically through tools, not free-form AI output.

High-Level Architecture

WhatsApp → Webhook → n8n → AI Agent (LangChain) → MCP Tools → External APIs / Database → Response

n8n acts as the orchestration layer, while the AI agent handles reasoning and intent classification. All real-world actions are delegated to MCP tools with strict contracts.

Placeholder: Detailed flow diagram or n8n workflow screenshot

Identity & Security Model

Before performing any action, Jarvis normalizes the incoming WhatsApp identifier and queries Supabase to retrieve the full user profile. User identity and the current date are injected into the system prompt, ensuring that the assistant always knows who it is acting for.

Identity resolved before execution
Per-user context isolation
No cross-user data exposure
Secure credential handling

Engineering Highlights

Failure-aware design

Workflows are built assuming partial failures, with explicit error paths and safe retries.

Deterministic execution via MCP

The AI agent never performs side effects directly. All actions are executed through MCP tools.

Conversational memory with system boundaries

Context is preserved for conversation, while execution remains auditable and deterministic.

Cost and latency optimization

GPT-4.0-Nano is used to balance responsiveness and operational cost.

OAuth-based calendar access

Google Calendar operations use access and refresh tokens obtained during user login.

System Maturity

Jarvis is designed as a production-grade system, handling real user data and real API integrations. New capabilities can be added by introducing new MCP tools without rewriting the agent logic.

Tech Stack

n8n (workflow orchestration)LangChain (AI agent)GPT-4.0-NanoMCP (Model Context Protocol)SupabaseWhatsApp APIGoogle Calendar API (OAuth)Docker / VPS infrastructure

"This project demonstrates how AI agents can safely interact with real-world systems when identity, execution boundaries, and observability are treated as first-class concerns."

Placeholder: Short demo video
Placeholder: Additional screenshots or logs