The lead research pipeline resonates. Hit the same noise problem when querying by company name — "Notion" matches hundreds of headlines about the concept of notion, not the app. Fixed it by combining domain + company name in the query string.

Also found that parsing script/meta tags directly from the homepage beats any third-party data source for tech stack detection. HubSpot, Salesforce, Stripe, Intercom all leave distinctive fingerprints in page source. Zero API calls, zero cost.

Built something similar for B2B prospecting (batch mode — 50 companies at once). Ended up with almost the same architecture: HTTP scraping at 0 LLM tokens, LLM only for the synthesis step at the end. The bottleneck is rate limiting from the target sites, not the LLM.

One thing I'd add: on the engagement loop bug you mentioned — I ran into the same thing early on. The fix was processing only items where a "last_engaged" timestamp was >N hours ago before feeding to the agent. Simple filter, saved a lot of wasted runs.