What talks to what
Architecture · end-to-end
Everything below runs inside Aspyre's infrastructure except the outbound call to Claude (amber, right).
What we're building the agent to do
Six jobs, one output — a ServiceTrade draft quote.
The agent runs per deficiency, stateless, event-driven. Each run produces a ServiceTrade draft quote with a confidence score attached. Office reviewers open the draft natively in ServiceTrade — no second UI, no second login. The six jobs below are the sub-capabilities the "Agent worker" row in the roadmap is actually delivering week by week.
1
Extract structured items
Parse the deficiency description. Pull out component, quantity, location, condition. Handle vague write-ups by dropping confidence — never invent details that aren't in the source text.
2
Match against the price book
Look up each item in Aspyre's price book (source of truth maintained by Aspyre). Apply current markup policy. If an item isn't in the book, flag for manual pricing — never invent a SKU.
3
Estimate labor
Apply labor-time heuristics learned from 1,987 historical accepted quotes. Distinguish standard vs. after-hours crew. Round to Aspyre's standard labor increments.
4
Scope language from NFPA
Reference the NFPA code library (Aspyre-provided) to write scope text that cites the correct standard — e.g., "Replace (6) water gauges exceeding 200 PSI per NFPA 25."
5
Validate against comparable quotes
Compare the draft to recent accepted quotes for the same service line in the same metro. Flag outliers — too high or too low — before a reviewer ever sees the draft.
6
Write the draft back for review
Produce a structured draft ready for office review. Current plan is a native ServiceTrade draft quote — if the workflow calls for something different, we'll build a small custom review surface instead. Decision in W2.