Inspiration
The rise of autonomous AI agents has revolutionized how tasks are performed—but most AI applications still rely on isolated functionalities. We wanted to explore how a coordinated team of specialized AI agents could collaborate to create high-quality, multi-format content (text, code, image prompts) from just a single topic. Our goal: Build a smart, end-to-end content generation pipeline that automates research, writing, refinement, coding, and visualization with minimal human input.
What it does
ContentCrafter AI is a multi-agent system that takes a single topic as input and produces:
A structured content plan
In-depth research insights
Long-form, informative content
Refined, well-formatted output (tone/language customization)
Python code examples related to the topic
Descriptive image prompts (for tools like DALL·E or Midjourney)
All of this is powered by six collaborating agents:
Planner Agent – Outlines the strategy
Research Agent – Gathers factual data
Content Generator – Writes long-form content
Refinement Agent – Polishes tone, clarity, grammar
Code Agent – Adds sample Python code
Image Agent – Crafts AI art prompts
How we built it
Framework: Python with ADK structure
UI: Streamlit for easy interaction
LLM Backend: GROQ API with Mixtral (and fallback to LLaMA 3 70B when Mixtral was deprecated)
Modules: Modular Python files per agent (planner, research, etc.)
Orchestration: A central OrchestrationAgent manages the flow from topic input to output delivery
Dotenv: For managing API keys securely
Challenges we ran into
🧠 GROQ Model Deprecation: Our initial model (Mixtral) was deprecated mid-development, so we reworked the system to support fallback models.
🔄 Agent Coordination: Designing agents to pass meaningful context between each other without redundancy or hallucination required multiple prompt iterations.
💡 Pydantic Modeling Errors: ADK’s BaseModel conflicts caused initialization issues, which were resolved by restructuring constructors.
🧪 Testing Streamlit Locally: Some environments failed due to API limits or misconfigurations, so fallback logging and verbose output were added for debugging.
Accomplishments that we're proud of
Created a functioning multi-agent pipeline with fully modular agents
Successfully integrated GROQ API into a custom LLM orchestration flow
Designed a clean, intuitive UI that makes advanced multi-agent tech usable for everyday users
Adapted to last-minute model deprecation and kept the project working
What we learned
How to build agentic systems using Pydantic, Streamlit, and modular design
Fine-tuning prompts for agents to reduce hallucinations and increase accuracy
Deep understanding of GROQ API limitations, model behavior, and fallback planning
How to manage LLM-driven workflows in an extensible and maintainable way
What's next for ContentCrafter AI
Add a Blogger publishing agent to auto-publish refined content
Integrate voice-over or narration generation for accessibility
Enable visualization support, turning data into charts/images automatically
Create multi-language pipelines using translation agents
Package as a SaaS microservice for content marketers, educators, and bloggers
Log in or sign up for Devpost to join the conversation.