№ 530305July 2026 · San Francisco, CA
Recruiting Desk
A system to find the best startup jobs for Amherst alumni.
Problem
Startup roles come and go by the hundreds, and Amherst's alumni network is full of people who could fill them beautifully, but nobody has time to read every role, remember every person, and notice when someone changes jobs. Matching by hand means the right candidate gets skipped quietly.
Solution
The desk keeps a live list of open roles, builds a ranked shortlist of alumni for each one with a plain reason and a verdict per requirement, and moves each person across a tracker board from first message to submission. It runs on the HerdSearch directory and its search engine.
Frequently finding great roles
Every day the syste, reads open roles and decides which ones are worth working.
Open roles arrive daily from an external source into an inbox, each with a suggested title family and seniority band, and a person marks every one in or out of scope. An in-scope role carries a fingerprint of the fields the search depends on, and its detail is refetched only when that fingerprint changes. Changes to requirements, salary and competing candidates are stored as events, and three rules turn them into alerts: a requirement changed, so re-screen; a rival's candidate reached a late round, so stop spending outreach; the salary band moved, so the pitch changes. Nothing is deleted; a role that leaves the feed just flips to inactive.
A matching system for the best candidates
With 10,000 people, AI cannot match people well. This had to be a deterministic mapping system handled for decent scale.
This is a search problem, then a sort problem. Search, because a role’s requirements must become three to six retrieval queries over an alumni database where most fields are unknown rather than false; hybrid search (vectors, BM25, full text) runs each, reciprocal rank fusion merges them into about seventy-five finalists, and code computes dated stints, years against the band and metro fit. Sort, because the desk is paid on interviews and acceptances, so the order is an expected-value product: P(engage) from a level meter, times the role’s real interview rate shrunk toward a prior, times proven fit, where an unknown discounts a little and a hard miss a lot. One function orders every queue, so a card never slides twice.