At Barracuda, we’re continually innovating to remain forward of rising safety threats in an more and more complicated digital panorama. As an organization trusted by a whole lot of hundreds of companies worldwide to guard their e-mail, networks, purposes, and knowledge, we perceive the essential significance of complete safety options. Barracuda exists to guard and assist prospects for all times – how can we leverage cutting-edge AI expertise to additional our mission?
As Principal Engineer main the Barracuda GenAI platform initiative, I understand how vital it’s to supply product groups with a consolidated regional, scalable, and compliant platform with minimal overhead whereas enabling them to confidently construct, iterate, and deploy AI options. Barracuda AI supplies easy accessibility to over 20 AI fashions, with assist for the most recent fashions added inside days by means of steady APIs. We depend on Databricks’ superior tracing capabilities to watch, troubleshoot, and enhance our AI platform and are actively engaged on integrating Databricks’ LLMOps options, corresponding to LLM Choose Metrics and Monitoring, to simplify LLMOps for product groups utilizing Barracuda AI.
Energy of Tracing for Barracuda AI
In cybersecurity, understanding precisely how AI fashions make choices is essential for each effectiveness and belief. Tracing supplies unprecedented visibility into our AI purposes, permitting us to trace each step of the decision-making course of from preliminary request to remaining response.
Once we noticed MLflow LangChain autologging at Databricks Knowledge + AI Summit, we built-in simply and have been benefiting ever since.
Tracing permits us to:
- Comply with the entire journey of a request by means of our system
- Determine bottlenecks and efficiency points in real-time
- Debug complicated interactions between a number of AI elements
- Guarantee constant habits throughout completely different environments
- Present audit trails for safety and compliance functions
By implementing complete tracing throughout our platform, we will rapidly determine and resolve points, optimize efficiency, and guarantee our safety options are performing at their finest whilst assault patterns evolve.
Our Technical Implementation
Barracuda AI is constructed on a basis of versatile, interoperable applied sciences designed to maximise efficiency whereas minimizing overhead.
Barracuda AI API Infrastructure
Our API provides OpenAI-compatible and LangChain AIMessage/AIMessageChunk endpoints (with extra coming quickly) that allow seamless integration with present instruments and workflows. This compatibility layer permits product groups to iterate and experiment with out worrying about deployments or code modifications throughout mannequin or agentic frameworks. Behind the scenes, we fastidiously wrap interfaces and deal with translations by means of a regional, scalable API gateway deployed by way of Kubernetes clusters and constructed utilizing FastAPI served by Uvicorn, making certain constant habits and efficiency whereas sustaining detailed tracing.
Barracuda AI Frontend
Barracuda AI additionally has a safe, SSO-authenticated Subsequent.js front-end software for wider AI utilization throughout the corporate.
Monitoring and Logging
MLflow autologging capabilities routinely monitor all mannequin interactions with out requiring in depth code modifications. This “set it and overlook it” strategy to tracing ensures we seize complete knowledge whilst our platform evolves.
Knowledge Processing and Evaluation
Databricks integration provides highly effective analytics and monitoring capabilities that enable us to course of huge quantities of hint knowledge effectively. For latest traces (inside the final hour), we use the MLflow UI for quick evaluation. For older exported traces, we’ve constructed views with DBT for our Databricks Genie area, permitting us to extract significant insights and analytics utilizing pure language.
Day-to-Day Utilization Eventualities
Our tracing infrastructure helps quite a lot of essential use circumstances that assist us keep safety excellence:
Troubleshooting Complicated Points
When customers report uncommon habits, our builders can instantly search for the related request_id and retrieve the corresponding hint. This permits them to hint your entire journey of that request by means of our system, figuring out precisely the place issues went unsuitable.
Complete Efficiency Monitoring
We have constructed refined dashboards and each day reviews that give us visibility into:
- Utilization patterns by crew and mannequin
- Price evaluation and optimization alternatives
- Token utilization monitoring for effectivity
- Mannequin efficiency metrics and latency statistics
These dashboards enable us to make data-driven choices about useful resource allocation and determine alternatives for optimization.
Abuse Detection and Prevention
Safety is about defending in opposition to each exterior threats and potential inner vulnerabilities. Our tracing system helps determine misuse situations, corresponding to when growth keys are unintentionally deployed in manufacturing environments.
Managing Giant-Scale Knowledge
Dealing with hint knowledge at scale presents distinctive challenges. For very giant traces containing huge context masses (corresponding to in depth code bases or giant copies of logs), we have carried out clever truncation methods to remain inside the 16MB JSON restrict of Databricks’ VARIANT sort whereas preserving essentially the most essential info.
We additionally prioritize knowledge privateness. For traces at relaxation in Delta Lake Tables, we take away personally identifiable info (PII) for knowledge safety functions whereas preserving the analytical worth of our hint knowledge.
Future Instructions
We’re actively exploring a number of thrilling enhancements to our Barracuda AI platform:
Superior Analysis Capabilities
Utilizing analysis and monitoring APIs is excessive on our precedence listing and on our hackathon roadmap. We plan to reveal these analysis capabilities by means of our platform APIs, permitting groups to measure and enhance the standard of their AI-powered safety options.
Democratized Knowledge Entry
Use Databricks Delta Sharing to permit groups to run their very own analyses on hint knowledge. This functionality will empower them to derive insights and drive modifications particular to their purposes.
Enhanced Offline Analysis
We’re growing capabilities for offline analysis of hint knowledge, enabling groups to check hypotheses and enhancements with out impacting manufacturing techniques. This strategy accelerates innovation whereas sustaining the soundness of our safety infrastructure.
Expanded Monitoring
As we incorporate new options and enhancements in our GenAI platform, we’re exploring methods to reinforce our monitoring capabilities. We need to speed up product innovation, like deploying AI brokers on Databricks that combine with our GenAI platform, and broaden the visibility of our tracing infrastructure.
Conclusion
Barracuda AI is a basis for future innovation at Barracuda, giving product groups the flexibleness, energy, and visibility they should construct the following technology of safety options. By centralizing AI capabilities, streamlining observability by means of tracing, and harnessing the scalable infrastructure supplied by Databricks, Barracuda AI has turn out to be a cornerstone that empowers lots of our product initiatives. Because the menace panorama evolves, we stay dedicated to defending prospects for all times by frequently refining and increasing this AI basis, making certain each Barracuda resolution advantages from strong, agile, and future-ready innovation.
