All work

Lead Engine

2026 · mhmudsaed/hermes-lead-engine

Built
Browser automationPythonxlsx

Takes a list of companies, visits each site with a real browser, and fills in the spreadsheet: hiring contact, tech stack, careers page — with a screenshot saved as the receipt for every row.

Built and verified · recorded run, offline test suite, and full source below.

The problem

A raw list of company names tells you nothing about who is worth writing to. Finding the stack, the hiring page, and a real contact by hand runs to a couple of minutes per company, which is hours for a list of fifty.

Pipeline architecture

The tool reads the input CSV and drives a real browser through each company's site in turn: homepage, careers page, about page. It collects what it finds into structured columns, drafts a two-sentence outreach note per row, and writes an enriched spreadsheet. Each enriched row stores the screenshot and source URL it was based on.

Recorded run

Recorded live in a real terminal: the offline test suite, then ./scripts/demo.sh against the synthetic fixture site — the enriched table and run report at the end.

The same enrichment with the browser window on screen: Chromium stepping through each company's pages.
pytest — 35 passed, including real-Chromium end-to-end tests
pytest — 35 passed, including real-Chromium end-to-end tests
Enriched rows: careers URL, published contact, tech signals, screenshot receipt
Enriched rows: careers URL, published contact, tech signals, screenshot receipt
The driven browser on a company page (synthetic fixture site)
The driven browser on a company page (synthetic fixture site)

Key decisions

Accessibility snapshots instead of raw HTML

Reading a page's accessibility tree instead of its HTML gives the browser agent clean element labels at a fraction of the tokens, and it holds up on the form-heavy career pages where raw parsing breaks.

One browser session per source site

Sessions stay isolated per site so cookies and login state never bleed between companies, and a problem on one site cannot take down the whole run.

Screenshots as receipts

Every enriched field keeps the screenshot it came from. If a number or a name looks wrong, the original page is one glance away instead of a re-browse.

Guardrails

  • Stops and asks when a site presents a CAPTCHA or a login wall.
  • Rate-limited, jittered browsing; it behaves like a slow human reader.
  • Fields without a verified source are left empty rather than guessed.

Verification

Targets turning a couple of hours of manual research into one run. Verified in the recorded runs below: the sample list enriches end-to-end with screenshot receipts, and the one CAPTCHA-walled page is skipped rather than guessed.