What is Martian? Martian builds model routers — systems that dynamically select the optimal AI model for each query. This approach allows businesses to leverage multiple LLMs, optimizing for performance, cost, uptime, and other business requirements.
We build these routers by doing research into interpretability . To know what model we should use on a given query, we have to predict how different models will behave on that query. Only then can we know the best model to use without having to run them first.
Martian & Accenture's Partnership We're partnering with Accenture in two key ways:
Powering Accenture's AI "Switchboard": Our technology will underlie Accenture's central orchestration layer for AI deployments. With Accenture's recent announcement of $1 billion in booked Generative AI revenue, Martian's router will play a crucial role in these initiatives.Long-term collaboration: Accenture is investing in Martian and including us in their Project Spotlight program where they feature their top partners. We share a vision of a multi-LLM future, where sophisticated applications will require orchestration across many models.We’re excited about working with Accenture: with their knowledge of enterprises and our technology, we can bring the cutting edge of AI to production across entire industries. For enterprises, Accenture is great as a neutral third-party — because they’re not tied to a particular LLM, they can use our technology to help companies adopt AI in an unbiased and optimal way.
Airlock: Making AI Adoption Easier Alongside the Accenture partnership, we're launching Airlock, our LLM Compliance Automation suite.
The Problem: Integrating new LLMs in enterprises is often slow and costly due to extensive legal and safety checks. This compliance burden can significantly delay AI adoption.
The Solution: Airlock Customizable Policies: Airlock can automatically read and extract the key requirements from your organizations policies (e.g. filter models to be GDPR compliant, have 0-day retention, or meet other legal requirements). You can then edit and add additional requirements for various teams or application types.Automated Testing : Every time a new model is released, Airlock automatically vets that model to see if it passes the full set of policies you have defined.Enforced Compliance : Approved model usage can be enforced directly in code.This approach maintains necessary compliance standards while accelerating AI integration timelines. As soon as a new model is released, you know whether you can use it.
We’re especially excited about this tool as an extension of our research into interpretability. When creating tests for models, we can use techniques from interpretability to look inside of these models in order to identify and mitigate unsafe or undesirable behaviors. This provides a level of safety above and beyond what is otherwise possible.
The Future of Enterprise AI Recent discussions in the AI community, including insights from Reid Hoffman , Andrej Karpathy , and Delip Rao , point towards a future of multiple, specialized AI models rather than a single, all-purpose solution. As models further diverge in their costs and capabilities (OpenAI’s recent o1-preview is 100x more expensive than the cheapest gpt-4o model, but gpt-4o still has capabilities such as multimodality or chat which are stronger than o1) and new models and providers enter the space (Poolside raised $400M+ and their flagship model still awaits release , Mistral recently released pixtral ), how you use AI will become as important as whether you use AI.
Martian and Accenture's partnership aims to make this multi-model future accessible to enterprises. By removing typical integration hurdles, we're enabling businesses to explore and utilize a wider range of AI models efficiently.
Our goal is simple: make it easier for enterprises to adopt and benefit from the rapidly evolving landscape of AI models. If you're interested in streamlining your AI strategy, sign up to learn more about Martian and Airlock.
