A neuro-symbolic prototype combining LLM-powered task decomposition with constraint-based optimization for intelligent scheduling.
Built for the Hugging Face Agents MCP Hackathon.
What it does#
Yuga Planner transforms project descriptions into optimized employee schedules:
LLM Task Decomposition
Constraint Optimization
MCP Integration
How it works#
Project Input
Markdown Parsing
Accepts project descriptions in markdown format with automatic task extraction.Task Decomposition
LlamaIndex + Nebius AI
Breaks down projects into actionable tasks, analyzing skill requirements and dependencies.Optimization
Timefold Solver
Generates optimal assignments respecting calendar constraints, business hours (9:00-18:00), and weekends.
Architecture#
sequenceDiagram
actor User
participant LLM as LlamaIndex
participant Solver as Timefold
participant Cal as Calendar
User->>LLM: Project description
LLM->>LLM: Extract tasks
LLM->>Solver: Task constraints
Solver->>Cal: Check availability
Cal-->>Solver: Free slots
Solver-->>User: Optimized schedule
Features#
Dual-mode operation: Works as both a Gradio web interface and an MCP tool for integration with agent platforms like Claude Desktop.
- Calendar integration with
.icsfile support - Real-time log streaming and progress indicators
- Streaming tool call processing with JSON repair
- Intelligent scheduling request detection
Tech Stack#
Python 3.10+
Java 17+
LlamaIndex
Timefold