AI Observability and Security for Government
Fiddler is a pioneer in AI Observability and Security, the foundation to ensure the performance, behavior, and safety of agents, predictive, generative AI models and applications. We partner with government agencies via a proven AI Observability and Security platform to achieve responsible AI.
We help government agencies to:
- Monitor agents, generative and predictive models and applications to ensure accuracy, safety, and privacy.
- Explain model and agent outcomes, and diagnose the root cause of their behavior.
- Protect LLM applications using real-time Guardrails.
- Enable human-in-the-loop for decision-making in mission-critical applications.
[00:00:00]
[00:00:01] As experts in AI Observability and Security, we see federal agencies rapidly adopting AI. But it comes with significant risks that must be managed.
[00:00:09] First, predictive models can be a black box. For example, a computer vision model might classify a sonar image incorrectly by focusing on an object's shadow instead of the object itself.
[00:00:20] Second, foundation models can hallucinate, confidently making up assertions that have serious ramifications for mission-critical decisions.
[00:00:27] And third, AI can introduce discriminatory bias in sensitive areas like benefits allocation or threat assessment. For federal agencies; these missteps can impact citizen wellbeing and national security.
[00:00:40] This level of oversight is crucial for high-stakes federal use cases, from fraud detection and benefits allocation to national security.
[00:00:47] Fiddler addresses four critical strategic focus areas: cybersecurity; AI/ML scaffolding; frontier AI; and situational awareness.
[00:00:56] Our work on the Project AMMO prototype for the Department of War advanced to production with NIWC, cutting US Navy ML model update time by 97%. This earned us a Defense Innovation Unit Success Memo.
[00:01:09] We've received the DoD app FIT award. we're TRL 7 production-ready, and also deploying into IL6 environments.
[00:01:16] The unified Fiddler AI Observability and Security platform serves as your Command Center for predictive, generative and agentic applications.
[00:01:23] We offer real-time alerts, interactive root cause analysis, customizable dashboards, and reports for audit trails. We monitor 30+ out-of-the-box ML metrics and over 80 LLM metrics powered by our native Fiddler Trust Models, and use real-time Guardrails to block harmful outputs and prompt injections with less than 100 milliseconds latency.
[00:01:45] Fiddler provides end-to-end agentic observability, giving you complete hierarchical visibility and root cause analysis from the application down to the span level.
[00:01:54] Now, let's see Fiddler in action...
[00:01:56] Fiddler is your AI Command Center, providing a comprehensive AI inventory to track all models across your enterprise. This empowers you to deploy AI quickly, safely, and effectively.
[00:02:07] Let's explore three key AI categories we monitor:
[00:02:10] First, Predictive Models, like this credit approval system. We track its accuracy over time. For fairness, we monitor average loan probability across demographics and identify bias using metrics such as disparate impact across gender or race. When issues like drift occur, Fiddler's root cause analysis immediately pinpoints where the data or prediction drift is happening. For instance, here we see a spike in low approval rates. Our explainability features then open the model's black box. For this specific credit request with a low approval probability, we can see that a high number of past dues and low paid-off balances were the primary negative features contributing to the model's decision.
[00:02:51] Next, Generative AI applications. We help you govern systems, like this chatbot, by evaluating their inputs and outputs in real-time. A common practice is "LLM-as-a-Judge," which uses one large language model to evaluate the responses of another. While other platforms require external API calls that move data outside secure boundaries for this, Fiddler provides this capability entirely within your environment, ensuring complete data security and compliance.
[00:03:19] Our real-time Guardrails are powered by Fiddler Trust Models. This is the engine that detects threats like hallucination and jailbreak attempts without hidden costs or data exposure. Watch as we instantly detect and block a potential jailbreak attempt based on a high-risk score from our Trust Models. All these interactions feed into our observability platform where we track metrics like faithfulness and cyber risk profiles. We also visualize these in a 3D view for deeper analysis, allowing us to rapidly identify problematic clusters of prompts and responses. For full transparency and compliance, every interaction has a complete audit trail, including source documentation and guardrail metric scores.
[00:03:59] Finally, for complex Agentic AI Systems, where individual agents, and their interactions occur in series or parallel. Fiddler provides complete visibility. Our interactive root cause analysis lets you isolate specific traces and spans. This means you can pinpoint exactly where and why issues occurred at every level of your agentic system. These detailed analyses are then rolled up into a visual and graphical view for faster mean time to resolution.
[00:04:27] From predictive models, to generative AI, to the most complex agentic systems, Fiddler is the AI Command Center that gives government agencies the oversight and audit trails they need to deliver high-performance AI, safeguard public trust, and maximize mission effectiveness and accountability.
[00:04:45]
