InkreaLLM

Since the emergence of multiagent AI systems, Data Scientists, AI Engineers, and Developers are working to find appropriate use cases where these LLM agents can bring value without compromising the quality and robustness of existing processes. One domain where these systems can help is creative writing, where occasional hallucinations or misunderstandings can sometimes spark new narrative lines or unexpectedly transform the entire story concept.

Inspired by the potential of collaborative creativity and the opportunity for a family "teambuilding" project (with my wife as my teammate), I quickly agreed to participate in the "LLM Agents MOOC Hackathon" hosted by UC Berkeley in conjunction with the LLM Agents MOOC.

The issue we decided to tackle was: "How could a multiagent AI assist someone in the writing process?" While using a single LLM chatbot like ChatGPT, Claude, or Gemini is relatively straightforward for writing, the complexity appears when creating longer works with consistent character development across chapters or scenes. Our potential writing formats could range from short stories and essays to novels, poems, or improvised dialogues.

Our multiagent system will take care of the writing's structure, characters, and narrative flow, adapting to the specific genre and type of writing. By carefully analyzing the user's answers and suggestions, it will define and create the text, fragment by fragment, always under the author's supervision and guidance. Unlike many multiagent AI systems that are doing a great work in process automation, we prioritize empowering the user to enjoy the creative process rather than struggle in structuring, searching or correcting the writing.

In the same time the user can suggest modifications to the structure or to the created text, and one of the agents will readily pick it and redesign or rewrite in no time (actually 2-4 seconds from our tests). Also the writing can be not only in English, but in any other language supported by LLM.

Our tech stack contains Python, LangGraph, LangChain, Streamlit, and OpenAI + Gemini as LLM companions. The multiagent AI system is designed as a graph, containing a number of subgraphs with agents designed to help each within a small part of the process. This approach potentially enhances answer quality, albeit with slightly increased latency—a trade-off deemed acceptable given that writing isn't a process where speed is paramount.

The git repo: here