Architecture Playground
Reference systems, in detail.
Five architectures I use in real engagements — each with its nodes, edges, guiding principles, and the trade-offs I've learned the hard way.
LLM copilot layered over Genesys Cloud
AI-Powered Customer Support
A contact-center architecture where every voice and digital interaction is transcribed in real time, enriched with intent + sentiment, and augmented by an LLM copilot that helps the agent without ever taking the wheel.
Nodes
Customer
Voice · chat · email · messaging
Genesys Cloud
Omnichannel routing + orchestration
AudioHook
Live audio stream
ASR + Intent
Whisper / Deepgram + classifier
LLM Copilot
Summarize · draft · recommend
Vector KB
Grounded, freshness-scored
Agent Desktop
Side-panel suggestions
Flow
- CustomerGenesys Cloudinteraction
- Genesys CloudAudioHookstream
- AudioHookASR + Intent
- ASR + IntentLLM Copilotcontext
- Vector KBLLM Copilotretrieval
- LLM CopilotAgent Desktopsuggestions
- Genesys CloudAgent Desktop
Principles
- Human stays in the loop — copilot suggests, agent decides
- PII redaction happens before anything leaves the transcript service
- Every suggestion is scored; low-confidence outputs never render
Trade-offs
- Latency budget forces small models on the hot path, larger models offline
- Vector KB freshness vs. cost — we re-embed nightly, not on write
- Copilot adoption depends on UI design more than model quality