How I built this.
The posting says to expect a walk through a workflow I built, and where I drew the line between AI and my own judgment. Here is that walk, for this site.
The workflow
MePick the targetDeepgram’s newest posting. I had it read in full before anything else was planned.
AIRead everything publicTwo research agents: one on Deepgram, one on eight rivals. 410 documentation pages and 231 changelog entries, searched.
AICheck each findingEvery Deepgram claim was checked against the saved page text. Nine reported contradictions became eight.
MeMake the callsWhat the plan leads with, what it cuts, and what stays off a public page.
AIBuildThe plan, the prototype and a private findings report, with me reviewing each.
The lead agent was Claude. It ran the two research agents, did the checking and wrote the pages. One working day, November 22, 2026.
The rules the agents worked under
- Every claim carries its source.A page address and where on the page. The page text is saved, so any claim can be checked again.
- “Not found” is never rewritten as “missing”.The agents had to say where they looked. Only a statement by Deepgram itself counts as confirmed absent.
- Read only.No account, no login, no form, no testing of any system.
- Separate agents for Deepgram and for rivals.So what one vendor says about another never stood in for that vendor’s own page.
Where the AI was wrong or thin
- One contradiction did not survive checking.A change in the customer count between two announcements has an innocent explanation. It was dropped.
- A third-party tutorial contradicted Deepgram’s own docs.The research agent flagged it as unreliable, and nothing from it was used.
- Two directories credit Deepgram with certifications it does not claim.Not used. Deepgram’s own pages are the source for what Deepgram holds.
- The rival rows are thinner than the Deepgram row.They were not checked line by line, and the evidence page says so.
Where my judgment came in
The agents can read 600 pages in an hour. They cannot decide what a hiring team should be shown.
- Read the company as its buyer would, before pitching anything.A plan that opens with “build a trust center” is a guess. One that opens with “buyers cannot see what you have” is a finding.
- Report what turned up.If I find holes, I bring them up. The findings went into a report.
- Keep that report off the public site.Some of it looks unannounced. It goes to Deepgram directly.
- Say what I cannot see.The plan has a section on what would change it, because I have no view of the pipeline.
- Stand behind the output.I read every page of this site before sending it. If something here is wrong, that is mine, not the agent’s.