Next Good Lead: source-backed B2B prospecting
A B2B prospecting product that treats model research as proposals and controls publication through captured evidence, review, and code checks.
From an offer to a reviewable prospect list
I created Next Good Lead from idea through implementation, using AI assistance. I own the product direction, integrated application, research workflow, and public launch site.
For a team shipping an AI workflow, there are two distinct problems here: deciding which proposed results may reach users, and keeping a failed call from discarding unrelated approved work.
The application takes an offer and a chosen market through candidate-company discovery, source capture, proposed fit notes, and outreach drafts. Users can inspect sources and edit proposed leads before preparing or exporting a draft.
Give models and code different responsibilities
Models search for companies and propose structured research. Application code fetches public pages and retains the source captures. A separate call to the same model and provider reviews qualification, contact suitability, and contradictions.
- Model: discover companies
- Code: capture public sources
- Model: propose research
- Model: review proposals
- Code: publish or withhold
Before publication, code validates the review's structured decision and checks that quoted text and contact details appear in the captured sources. Missing required support blocks publication. The model review is not independent external validation, and the source checks do not prove that a source is correct or that a model interpreted it correctly.
Contain a correction failure
In a recorded run, a server error during an optional correction of one candidate stopped the whole workflow before a separately approved candidate could be published. The fix preserves approved candidates when that correction fails with a server error before any change is applied, and withholds the unresolved candidate.
| Test condition or assertion | Recorded result |
|---|---|
| Optional correction response | Injected HTTP 502 server error |
| Completed run | 1 lead published, the separately approved candidate |
| Usage accounting | 1 uncertain provider call; total cost estimate unavailable |
| Replay of the completed run | No extra provider calls or duplicate leads |
The fix adds no retry and does not relax the evidence checks. These are synthetic test results, not proof of live recovery; the original server error's cause remains unknown.
What remains to be demonstrated
Live trials have included a wrong deadline, an unsupported name citation, and a discovery failure before candidates were produced. Withholding unsupported claims does not establish that the product reliably finds useful opportunities. Research quality and provider reliability remain under evaluation.
The application keeps saved leads and sources within a workspace for review, editing, and draft export. Outreach remains under the user's control; the product does not send messages automatically.
The application is in internal testing. The public website accepts requests for a future preview notice; its illustrated product example is fictional.