1
It’s already running in production in a semi‑automatic way (every action still goes through human approval). So far the system has sent about 200 texts. Too early to measure impact, but it’s been enough to see interesting behavior.
=== How it’s built:
Almost stateless: core data lives in the CRM, with SQLite for caching / minimal state. Written in Typescript, using a lot of functional programming. Trying to follow Domain Driven Design patterns, which keep the logic and bounded contexts manageable. Rate limits and random API errors were a pain: solved with higher‑order functions wrapping API calls. Challenges so far:
Prompt tuning: lots of “close but not quite” AI decisions, especially edge cases where I’d act differently as a human. Infrastructure: significantly more code around state, retries, and API integration than around the actual LLM.
=== My question to you:
Does this sound like something worth pursuing further? Is there a genuine gap here, or am I just reinventing existing automation tools?
I’d love to hear perspectives from people who’ve built or used similar systems.