Practice Finder

An account-based outreach tool for Zenara Health. Give it a specialty and a US region. It searches Google Maps, reads each practice's own website, scores every practice against an ICP you define, and drafts an icebreaker from Dr. Ravi Hariprasad for the ones that clear your bar.

It stops at drafted. Nothing is sent automatically.

Pipeline

1. Source - SerpApi Google Maps engine, live query, paginated and deduped
2. Read site - Jina Reader fetches the homepage plus a linked about, team or contact page, and pulls out published contact addresses, named clinicians and care-model signals
3. Qualify - Claude scores six dimensions 0 to 5 against your ICP, citing evidence from the fetched pages, and flags hard disqualifiers
4. Draft - Claude writes a first-touch note that opens on a detail specific to that practice

Steps 2 to 4 run with bounded parallelism, so a batch of twelve takes about as long as a batch of four used to.

Defining the ICP

Everything the qualifier judges against is editable in the UI, and everything is passed through to the model on each run:

Field	Default
Specialty	Psychiatry
Region	any US metro you type
Clinician band	2 to 15
Independent only	on, which disqualifies hospital and health-system practices
Positive signals	outpatient psychiatry, private practice, accepting new patients
Disqualifying signals	hospital, health system, academic medical center, inpatient only, urgent care
Free-text notes	anything else that defines a good account

Scoring

The model scores six dimensions from the evidence it found. Those are combined with weights you control to give a score out of 100 and a tier from A to D.

Dimension	Default weight
Practice size	30
Independence	25
Specialty match	20
Region match	10
Care model fit	10
Reachability	5

Weights are applied in the browser, so changing them after a run re-ranks the list instantly without spending another API call. A practice only gets an email drafted if it clears the threshold you set, defaulting to 60.

Results can be filtered by verdict and exported to CSV, including the drafted subject line and body.

Environment variables

SERPAPI_API_KEY=...       # required, Google Maps sourcing
ANTHROPIC_API_KEY=...     # required, qualification and drafting
ANTHROPIC_MODEL=...       # optional, defaults to claude-opus-5
ANTHROPIC_WORKSPACE_ID=... # only if the Anthropic key is not workspace-scoped
JINA_API_KEY=...          # optional, raises the site-reading rate limit

See .env.example.

Run locally

npm install
npm run dev

Deploy

Push to a repo and import into Vercel, or vercel deploy. Set the two required environment variables in the Vercel project settings before the first run.

Note that the API routes are unauthenticated. Anyone who can reach a public deployment can spend your SerpApi and Anthropic credits, so put the deployment behind Vercel's password protection or an auth layer before sharing the URL.

Design notes

• Contact addresses are only surfaced where a practice publishes one openly on its own site. The agent never guesses or pattern-matches an address.
• The clinician list is a regex heuristic over the site text. It is handed to the qualifier as weak evidence and labelled as such in the prompt, so the model does not treat it as an authoritative headcount.
• Qualification prompts forbid inventing clinician counts, affiliations or any fact not present in the fetched page text.
• Practices whose evidence does not settle size or ownership are marked "pending" for a human look rather than being force-classified.
• Every draft has to state plainly that Flow does not diagnose, treat or replace clinical judgement.

On the sourcing API

SerpApi is used because it returns Google Maps results without a Places API contract and needs no billing setup. Reasonable alternatives if volume grows:

• Google Places API directly. Cheapest at scale and first-party, but it needs a billing account and its terms restrict how long results may be stored.
• Outscraper or Apify Google Maps scrapers. Better at bulk export and returning more fields per place, slower per call.
• Exa or Nimble for the reading step, in place of Jina, if you want search-driven enrichment for practices that have no website at all.

The sourcing step is one route with one response shape, so swapping it is a contained change.