Building Copilots - key lessons and best practices
summary
An always-on contact-center copilot that translates live calls in real time is built and running in production using Azure Cognitive speech-to-text, Azure AI fine-tuning, OpenAI voices, and GPT-4 Turbo (migrating to GPT‑4o); initial 20-second latencies drop to consistent responses after adopting PTUs for consistent latency. A template orchestration (session manager, planner, extensions like SQL/RAG), evaluation SDK and prompt flow tracing for CI/CD testing, PII redaction API, and telemetry-driven dashboards prove the system workable while tracking cost, quality, and user signals.
AI-written summary · about
chapters (20)
- 0:00Building effective Copilots
- 1:16Parloa's AI agents for contact centers
- 3:12Azure services for Copilot
- 4:57Testing AI agents at scale
- 6:25Future of AI agents in support
- 8:12Tools to build Copilots
- 9:38Design principles in Azure AI Studio
- 12:33Prototyping Copilots with Azure AI Studio
- 14:39Experimenting in agent architecture
- 15:42Demo: SQL analytics agent
- 20:35Prompt engineering and search
- 23:27Evaluation data investment
- 25:58Evaluating with Azure AI Studio
- 32:46Key evaluation lessons
- 35:41Monitoring and tuning in production
- 37:13System metrics and latency
- 40:25PII detection in production
- 41:29Optimizing GenAI latency
- 44:21Production architecture and costs
- 44:44Future of AI applications
related talks
46mBuild and maintain your company Copilot with Azure ML and GPT-4
45mThe era of the AI Copilot
51mCan AI Face Your (Potential) Customers?
31mBuild AI-assisted communication workflows for customer engagement
15mAI Studio - Creating and scaling your custom copilots
27mBuild a RAG-based Retail Copilot Code-First on Azure AI Foundry
42mUX: Designing for Copilot
45mWelcome to the world of SQL Copilots