Talent Engineering at Polymarket

Strategy

Understand who is great and why, inferred from how people show up in the world not only where they worked (though this is important and is a big part of the strategy). Signals include what they post and reply to, who they follow and who follows them, what they ship (repos, writing, design, markets), and any other signals that are relevant to the role. When combined this should reliably produce the insights you want for identifying the best people to reach out to.

Talent is segmented and not one-size-fits-all. It is not the same company lists, discovery rules, or bar for “great” across engineering, policy / government affairs, CX, design, or markets-facing roles. Each segment runs its own custom company lists (who counts as a strong source org, tiering, and who to pull from) and its own custom evals for judging talent. Rubrics, scorecards, and model prompts calibrated to what excellence means in that lane. Sources and enrichment depth follow from those lists and evals, not from a generic template.

Data sources

Pull structured and behavioral signal from APIs, scraped pages, public artifacts, and community surfaces scraping aggressively where it matters. The architecture map lists categories; the point is coverage + cost: no single source owns “truth,” and every ingestion path is budgeted.

Across the large platforms, the workhorse pattern is regex plus small LLMs: build a lean text blob per account or entity (username, bio, a handful of short posts or snippets), strip obvious junk with deterministic filters, then ask a cheap model for a tiny structured output — a label, score, or tag. That keeps token mass down and makes full passes over huge surfaces (including follower-scale lists) economically sane; see cost section for concrete token math.

Company intelligence

Build a living model of which organizations train the people we want: tiering, roster change, talent density, and narrative (what the company actually does vs headline). Those org sets are custom company lists per segment. This means that engineering’s A-list is not the same as policy’s or CX’s. The model feeds discovery seeds and tells you when a “B-tier” company is temporarily a better bet than a brand name for that specific lane.

Discovery

Expand from strong seeds: people and companies we already believe in through graph moves (social graph, co-attendance, co-contribution) and event-driven spikes (launches, posts, wins). Followers and following are a first-class volume surface: a seed account’s audience is a discovery frontier we can score in bulk (who matches the segment, who is noise) without hand-reviewing millions of rows.

I'd optimize for precision at the top of the funnel, not maximum raw leads. Regex gates and cheap model passes exist to shrink the set before we make expensive calls.

Identity resolution

One person, many handles: deterministic joins where possible (email, wallet, stable IDs), probabilistic joins with scores everywhere else.

Enrichment & scoring

Tier spend: cheap features for everyone in the frontier, then structured LLM extraction for maybes, deep passes only for people who clear a bar. The wide layers lean on regex pre-filters + small models with minimal output schema; the narrow layers can afford richer prompts. Custom evals per segment: scorecards, gates, and often different feature sets all trained on hiring-manager feedback so engineering, CX, policy, and other lanes are not sharing one generic rubric.

Role pipelines

Parallel tracks share infrastructure but not definitions of “great.” Each pipeline gets templates for sourcing, eval rubric, and which activation levers matter (e.g., competitions and OSS for eng; dinners and briefings for policy-aware roles).

Activation & events

Deliberate IRL and community moments: hosted comps, small dinners, curated invites. With attendance and follow-up feeding back into discovery and scoring. We'd work backwards from the info we have on people to find out what events they'd be interested in.

Outreach & feedback loops

Outbound that respects context (why you, why now), plus tight loops from recruiter and hiring-manager decisions into weights and seeds. Every “not a fit” would improve the next pass.