IntentMesh: Intent to Verified Action

System Architecture
Overview Dashboard & Natural Language Intent Creator
Interactive Visual Multi-Agent Flow Canvas
Deterministic Execution DAG (Plan View)
Granular Step Specification & Human Consent Policy
Human-in-the-Loop Safety Approval Centre
Live Workflow Execution Monitor HUD
Multi-Agent Groq LLM Workflow
Verified Outcome & Synthesized Results View

Inspiration

IntentMesh was designed to address the fragmentation caused by disconnected services. Our goal is to provide a unified platform that transforms natural-language requests into reliable, multi-step workflows while keeping users in control.

What it does

IntentMesh is a multi-agent workflow orchestration platform that converts user requests into structured, multi-step workflows. The Planner creates an execution plan, the Researcher gathers context from simulated Google Calendar and Gmail, the Executor performs approved actions, and the Verifier validates the outcomes. Users can review plans, approve or reject actions, monitor execution, and inspect workflow histories through an interactive web interface.

How we built it

We built the backend using FastAPI and exposed workflow capabilities through FastMCP with Streamable HTTP. The frontend uses React and Vite, while SQLite persists workflow states, execution results, approvals, and event logs. A shared orchestration engine powers both the REST API and MCP interface. We implemented idempotency keys to prevent duplicate write operations, persisted-state verification to validate execution outcomes, and deterministic fallback planning to keep the system functional when an LLM is unavailable.

Challenges we ran into

One of our main challenges was ensuring reliable execution when workflow steps failed or were retried. We addressed this through explicit workflow states, dependency management, approval checks, and idempotency keys to prevent duplicate operations. Another challenge was maintaining consistent behavior across REST and MCP interfaces. We solved this by routing both interfaces through the same orchestration engine and shared business logic.