For Engineers: Implementation Details & Stack
Engineers care about the concrete path from event to side effect. Default design is event in, policy decision, optional model call, CRM write, audit row. We keep the path short so failure domains stay obvious. Native CRM webhooks or polling adapters feed a queue. Workers execute with retries and dead-letter visibility.
Stack selections stay deliberate. CRM SDKs (Salesforce, HubSpot, Dynamics) own field-level security so we do not reinvent ACLs. A workflow engine coordinates multi-step jobs such as freight quoting packs where several external systems must agree. Models enter only when rules alone fail: classification, scoring, drafting. We pick smaller hosted models when latency and cost beat large general models for the task shape.
From prior delivery we reuse patterns. Quote automation needed orchestration across rate inputs and human gates. We structured that as discrete tasks with clear compensation if a later step failed. Property search ranking taught preference features that transfer to lead ranking. Both remain thin layers rather than new platforms staff must learn.
Optimization focuses on cache-friendly lookups, batched CRM writes, and throttling against known API limits. Idempotency keys stop double updates during retries. Shadow mode compares model output to human decisions for a week before cutover. Edge cases include partial payloads, daylight saving shifts on SLA clocks, and multi-currency quirks for Virginia exporters.
Observability is first-class. Structured logs carry correlation IDs from form submit to CRM update. Metrics cover queue depth, decision latency, model token cost, and write success ratio. Traces jump from gateway to worker to CRM call so on-call can isolate blame in minutes.
Local DX keeps teams fast. Infrastructure as code, containerized workers, and seedable sandboxes mean a new engineer can recreate the path without tribal knowledge. Pull request templates force security and cost notes. That discipline is what makes the build ownable after we leave.