Show HN: Architecture question: running an LLM as core infrastructure(automazionezeli.com)

1 pointsby senza1dio4 maanden geleden0 comments

Show HN: Architecture question: running an LLM as core infrastructure

https://automazionezeli.com

I've been experimenting with running an LLM not as a chatbot but as the core runtime of a business system, and I'm curious how others approach this.

The idea is that the model doesn't just answer questions but orchestrates tools and interacts with real application logic.

The architecture I'm currently testing includes:

Runtime

tool orchestration parallel tool execution loop detection circuit breaker / timeout guards token budgeting Context

context compression dynamic token ceiling Caching

deterministic LLM response cache semantic cache using pgvector Memory

short-term session memory longer-term semantic memory Evaluation

prompt evaluation set to test tool reasoning and failures I'm trying to figure out which parts are actually necessary in production and which ones are over-engineering.

For people building LLM systems beyond simple chat interfaces:

how do you handle tool orchestration? do you implement memory layers or just rely on context? are semantic caches worth it in practice? Curious to hear how others structure this.

0 comments

—