Skip to content
Allan Henslee, PhDResume (PDF)

Case 02A medical device client · 2026

AI research where every claim has a source

A research pipeline that writes outreach talking points, each with its source and a confidence level.

Year
2026
Client
A medical device client, through Robots & Bacon
Role
Designed and built the research pipeline and review tools
§ 01

The problem

A medical device startup needed to reach the right specialist physicians, with outreach that was personal and accurate, without a research team.

§ 02

What I built

A grounded-research pipeline.

  • Pulls physicians from public professional directories and the national provider registry (NPI, a public federal registry).
  • Researches each one from public sources and writes 1–3 ranked talking points per physician, each with its source link and a confidence level.
  • Drafts go to a review screen where a person edits and sends; replies are filed back automatically.
  • Also for the same client server-side A/B testing with an in-app results panel; an email-automation overhaul; a meeting-notes inbox that turns forwarded notes into a summary and an action-item table with owners and due dates.
§ 03

My part, and who else

The research pipeline and the tools above are mine. The client's existing site was built by partner Matt Rundle.

§ 04

How it works

  1. 01Public directories
  2. 02Registry matchNPI
  3. 03ResearchPublic sources only
  4. 04Source and confidenceEvery claim cited
  5. 05Human reviewA person edits and sends
  6. 06Send
  7. 07Reply filed
Fig. 1 Directory to filed reply; nothing is sent until a person approves it.

How I know it works

Tests, review, measured results

CheckHowResult

Hand-checked sample

Talking points from one pilot batch checked against their source papers

26 of 26

Correct

Citation check

Publications cited in the same batch; the one bad citation was caught and dropped before anything was sent

53 of 54

Cost per physician

AI processing per physician (estimate)

A few cents

§ 05

Decisions

  • A source link and a confidence level on every claim, so a reviewer can check any line in seconds.
  • A person sends every message; the pipeline only drafts.

Why it matters in healthcare

Traceability: every claim tied to its source, the same discipline as a design history file. The same pattern works for KOL mapping and market research.

§ 06

What changed

Personal, accurate outreach to specialist physicians without a research team.

Under the hood

1–3

talking points per physician

26 of 26

hand-checked points correct

  • Claude
  • Anthropic SDK
  • NPI registry