Businesses exploring generative AI need ways to connect language models with the information and tools they already use. Fixie is developing a platform for building agents that can link those models to company data, software and workflows.
Agents connect models to business tools
Fixie describes its service as a platform for creating language model experiences for enterprises. Its agents can work with databases, APIs, productivity tools and public data sources, then use the information they find to produce or manipulate text and images.
One example is customer support. An agent could receive a customer ticket, look up purchase history, issue a refund when appropriate and draft a reply. That workflow brings several tasks together through a natural language interface.
Fixie says agents can be written in any programming language and hosted on any infrastructure. Each can use a model selected for its purpose. The platform supports OpenAI's GPT-4, and customers can also bring their own models or choose other commercial and open models.
A platform focused on flexibility
Co-founder and CEO Matt Welsh argues that natural language can help separate systems communicate without requiring teams to write extensive integration code. In this approach, each system is wrapped in an agent interface, and a language model helps interpret and move information between them.
That idea overlaps with other efforts to connect AI systems to outside services. OpenAI's ChatGPT plugins offer a way to connect its model with external APIs, while Zapier's Natural Language Actions lets developers move information among apps and services using natural language.
Welsh says Fixie's distinction is that it is model- and provider-agnostic, and lets enterprises host agents on their own infrastructure. The platform also manages underlying model interactions and operational details such as user identity, authentication, sessions, storage and configuration.
Fine-tuning and the limits of AI
Fixie does not train its own language models from scratch. Instead, Welsh says customers can fine-tune existing models for their agents using proprietary information or historical data that passes through an agent. He presents that approach as a way to avoid the cost of training a model from the beginning.
The company also says it can constrain what models do, with the aim of making them more reliable at carrying out tasks and answering questions. Those controls do not resolve every weakness. Welsh acknowledged that models can still make up facts, a problem known as hallucination, and that Fixie does not solve issues such as bias or short memories.
Other fine-tuning options exist, including the open source LangChain and Llama Index. Fixie's stated goal is to make deployment accessible to people with different levels of technical expertise, while bringing model choice and infrastructure options into one platform.
Early access and backing
Fixie reported around 5,000 users in an early access program and said it was working with a wide range of companies on business automation, customer support, generative AI and graphics. The company planned a public launch in the coming days, with free personal use.
It announced $17 million in investment, comprising $12 million in seed funding and $5 million in pre-seed funding. The investors named were Redpoint Ventures, Madrona, Zetta Venture Partners, SignalFire, Bloomberg Beta and Kearny Jackson.
At the time of the announcement, Fixie had an eight-person team and Welsh said it planned to grow to 20 by the end of the year. The company said customer acquisition would be its main focus before then. Its broader proposition is that businesses can connect language models with existing systems, while retaining choices about models and where agents run.