Building a Personal AI Assistant: The Journey of NinjaTech AI and MyNinja.ai

At NinjaTech AI, our mission is to make everyone more productive by taking care of time-consuming complex tasks with fast and affordable artificial intelligence (AI) agents. We recently launched MyNinja.ai, one of the world’s first multi-agent personal AI assistants, to help drive towards our mission and provide users with a seamless experience.

MyNinja.ai is built from the ground up using specialized agents that are capable of completing tasks on your behalf, such as scheduling meetings, conducting deep web research, generating code, and assisting with writing. These agents can break down complicated, multi-step tasks into branched solutions and dynamically evaluate generated solutions while continuously learning from past experiences. All of these tasks are performed in a fully autonomous and asynchronous manner, allowing you to continue your day while Ninja works in the background and engages with you only when necessary.

One of the key challenges in building a personal AI assistant is ensuring that no single large language model (LLM) is perfect for every task. Thus, we needed multiple LLMs optimized for specific tasks that could work together in tandem. Additionally, we needed scalable and cost-effective methods for training these various models, which led us to utilize AWS Trainium chips.

In building our cutting-edge productivity agent NinjaLLM, we recognized the need to create a diverse dataset and fine-tune models for specific downstream tasks and personas. By using the Lima approach for fine-tuning, we were able to construct a supervised fine-tuning dataset and iteratively improve our models with the help of user feedback and performance benchmarks.

By utilizing Trainium chips, we were able to efficiently parallelize training and quickly iterate through fine-tuning rounds at a fraction of the cost compared to traditional training accelerators. The Neuron Distributed training libraries also enabled us to fine-tune and refine our models effectively, especially with the release of Meta’s Llama 3 models, allowing us to rapidly upgrade and prepare for launch.

For model evaluation, we used benchmark datasets such as HotPotQA and Natural Questions (NQ) Open, achieving notable accuracies with our enhanced Llama 3 RAG model. Looking ahead, we plan to further improve our model’s performance by using ORPO for fine-tuning and building a custom ensemble model from the various models we have fine-tuned thus far.

In conclusion, building next-gen AI agents to enhance productivity is at the core of NinjaTech AI’s mission. With the help of AWS’s AI chips, open-source models, and training architecture, we were able to create a groundbreaking personal AI assistant that empowers users to tackle tasks efficiently. To learn more about our journey in building NinjaTech AI’s multi-agent personal AI assistant, feel free to read our whitepaper or try out these AI agents for free at MyNinja.ai.