Inspiration
Pricing pages change quietly. A sales or product team needs to know what changed, but a screenshot diff creates noise and an AI summary can invent a price, billing period, or explanation. We wanted an alert whose numbers a person can inspect before acting on them.
What it does
PivotWire is a pricing-change agent for SaaS founders, product marketers, and sales teams. It compares stored page snapshots, identifies structured changes, and checks each required quote against the correct plan, value, and changed text. A user can open the before and after receipts, inspect the exact source span and snapshot hash, and review an evidence-based brief.
Our track is Live Wire. The working pipeline acquires a page through You.com, proposes structured pricing facts, verifies them, compares them with the last accepted baseline, and delivers a Slack alert for a supported change. It distinguishes price increases, free-tier removal, reduced limits, sales gates, seat minimums, and add-ons. The public dashboard also provides an explicitly labeled synthetic demonstration that visitors can explore without spending API credits.
How we built it
The backend uses Python, FastAPI, Pydantic, and HTTPX, with a JavaScript/HTML/CSS dashboard and Server-Sent Events for the activity trace. Stored snapshots and run records preserve the evidence behind each result. A persistent SQLite reservation ledger bounds provider spending.
You.com is the source and reasoning engine: Contents retrieves pricing pages; standard Research proposes structured pricing facts; scoped Search and Research support announcement investigation. Pricing proposals must pass deterministic checks against the stored Contents snapshot before they can become accepted facts. Research output is a proposal, not its own proof. The selected mode does not use Inference Router.
OpenCrawl coordinated the build: one session shared a cache-authentication finding, another consumed it and implemented six regression checks, and the build session consumed the result for integration. OpenCrawl also coordinated ownership of the live test and shared API budget. This is actual development coordination, separate from the product's runtime pricing pipeline.
Challenges we ran into
A plausible extraction can still attach a real quote to the wrong plan or infer an unsupported billing period. We kept strict plan/value/source binding and added one bounded correction attempt for evidence-invalid proposals. Failed or incomplete pages preserve the last good baseline. Authentication and credit failures cannot silently become cached success.
Accomplishments
The deployed acceptance test fetched a fresh nonce through You.com Contents, then detected a fictional sandbox Pro price change from $12 to $23. It produced six verified change types and nine exact source receipts, and delivered one explicitly labeled Slack alert. A duplicate check produced no additional alert. The saved run is available in the demo for inspection.
A separate live check extracted one current GitHub pricing snapshot after a bounded correction. That is single-snapshot evidence, not a claim that GitHub changed its prices. The public demo has also been exercised from an independent hosted runner and a phone outside the tailnet.
What we learned
Evidence verification is a separate job from generation. Finding a quote proves that text is present; it does not prove an inferred motive or the semantic truth of every research claim. Explicit provenance makes the boundary understandable to users.
What's next
Broaden real-company monitoring, evaluate extraction quality across more pricing layouts, and move the temporary demo to durable hosting. The current public demo runs from a laptop through Tailscale Funnel and requires that laptop to remain awake and online. We are not claiming a historical real-company pricing change, continuous-monitoring reliability, or benchmark accuracy from the sandbox test.
