Fiddler AI Observability for Federal Agencies

Fiddler is a pioneer in AI observability, the foundation to ensure the performance, behavior, and safety of predictive and generative AI models and applications. We partner with federal agencies via a proven AI observability platform to achieve responsible AI.

We support federal agencies in seven strategic focus areas. Improving situational awareness, increasing safety, AI ML scaffolding, and AI assurance, streamlining business processes, increasing autonomy and mobility, assuring reliable data sources, and discovering blue sky applications.

The challenges that federal agencies face are time wasted in debugging model issues, costs related to development, deployment, and maintenance of models, ability to monitor the correctness, safety, and privacy of LLM models and applications, ability to monitor subtle changes in image models and determine the robustness of LLM prompts. Ability to scale and put more models into production.

As a result of these challenges, The White House Executive Order on Trustworthy AI requires all federal agencies to enable AI monitoring and transparency by August 2024.

Fiddler offers MLOps and LLMOps monitoring, analytics, explainability, and protection with a mission to build trust into AI through responsible governance. We help federal agencies to validate, evaluate, and ship tabular, natural language, computer vision, and large language models into production, explain model outcomes, and diagnose the root cause of their behavior. Monitor models to ensure high performance, safety, correctness, and privacy, and receive alerts as soon as issues come up. Enable human in the loop for decision making in mission critical applications.

Fiddler was issued a success memo and OTE by the Defense Innovation Unit for Project AMMO and is transitioned to production with NW IC Pacific in the Navy. The Fiddler AI Observability platform empowers model developers in Project AMMO to identify and act on common failure modes of their undersea threat detection models using semantic clustering and model explainability.

The platform can help monitor the model's degradation and understand how models see differences in tranches of training data and request only the highest impact additional training data, which can be resource intensive for the Navy to acquire. Fiddler also helps boat operators identify and understand issues that the deployed model can face and enhance decision making with visual explanations of model decisions.

Fiddler has also partnered with In-Q-Tel as a portfolio company.

Let's see Fiddler in action.

Say you're using Fiddler to monitor a classification model that identifies airplanes, ships, and an empty seabed in sonar imagery, and if you have a real time use case, you would receive an alert that there's been a shift in the distribution of images that your model is seeing.

Fiddler does this by tracking the distribution of embeddings, vector representations that encode how the model sees your data. So, you go into the Fiddler vector monitoring chart to diagnose the issue. Fiddler's vector monitoring uses a unique patented clustering based algorithm, which looks at how those embeddings populate the high dimensional semantic space to identify change over time.

What you see here is a shift in the input to the classification model.

A way to visualize the differences between baseline and production inputs is to plot them into a 3D UMAP. Points that are close are considered similar by the model. You can see this cluster is a seabed without any objects and all of the classified objects end up in a cluster over here.

You can also zoom into the cluster. Here is a group of production predictions that are a little bit off the distribution from the reference data, and it's probably the domain shift detected and triggered the alert. You can select this cluster to analyze further.

With image explainability, you can understand what is the model responding to and why it's making its prediction on these images. The model pays a lot of attention to acoustic shadows and not the actual ship, which indicates a vulnerability to uneven seabed types that cast a lot of shadows that may confuse the model to classify shadows as a ship.

You can adjust the threshold for the visualization across these images. You can also overlay the explainability analysis and view each image close up.

Now that you know what data inputs are causing the drift, you can monitor segments of the data like input ship images, or metadata like water temperature or geographic region, and be alerted as soon as those segments drift, affecting the computer vision model to behave differently.

Fiddler also enables you to monitor, analyze, and protect LLMs in generated AI models.

Say you are building an LLM chatbot for the U.S. Department of Homeland Security's Citizenship and Immigration Services to train officers.

You want to monitor LLM metrics like

PII, Hallucination, Toxicity, and User Feedback. You can also create and monitor custom metrics in the Fiddler AI Observability platform that are unique to your application. For example, you'd also want to control costs you'd incur on API calls that are made on both accurate and inaccurate responses, so you can calculate the true costs of using the chatbot for officer training.

Unlike other solutions and homegrown tools, we are the only solution that offers image explainability and deep model diagnostics to find the root cause of issues. Fiddler has dedicated data science expertise and offers white glove support to incorporate the latest AI techniques for your solution.

We care deeply about our customers success and build a long term partnership to support responsible AI.