# architecture

pii redaction can't be a progressive enhancement

llm router instead of just retrying

rate limits kept killing chat. built a router.

webhooks scoped to conversations, not projects

building a real-time conversation capture system

browser-based recording to structured transcripts, no app install

designing a transcription format for LLMs

every field you add has a token cost

auditing our event-driven architecture

the greenfield temptation

dramatiq is not your architecture

real-time transcription pipeline for in-person conversations

making ECHO model-agnostic with LiteLLM

three transcription migrations in three months

openai to azure to self-hosted whisper on runpod