Use cases
Most automation around LLM agents has the same shape: items come in, code and agents process them, people decide where it matters, and something goes out. Each page below shows one such pipeline as a graph and a pipeline.yaml, then what Millstream adds and which figures tell you it is working.
Pipelines
Sourcing and scoring. Crawl several sources daily, dedupe, score each listing with an agent, price the shortlist and have an analyst approve before a PDF report. This is Millstream's reference pipeline, and the one its self-improvement figures were measured on.
Lead qualification. Take inbound leads from a webhook, dedupe and enrich them, score the fit with an agent, and route the promising ones to a sales inbox before the CRM.
Document intake. Fetch invoices and contracts, extract their fields with an agent, validate them in code, and route the exceptions to people while sampled rules approve the clean ones.
Support triage. Classify tickets, draft replies with an agent, and have staff approve or edit each one, with urgent tickets in their own queue.
Content enrichment. Import a supplier feed nightly, skip unchanged products, fill in categories and attributes with an agent, check them against the taxonomy, and review what the rules cannot approve.