Skip to content

Professional networking

Reference build

Three times the industry connection rate, in 15 milliseconds

A matching engine that models people as vectors across skills, goals, seniority, industry and geography — and returns a ranked, business-rule-aware shortlist in fifteen milliseconds.

3xindustry connection rate

At a glance

Duration
39 weeks
Team size
4 people
Engagement
New build
Project type
Web application
Industry
B2B SaaS

The situation

The challenge

Good matches need signal, and asking users for signal directly means a preference form nobody finishes. Beyond that, naive pairwise matching is O(n²) and stops being computable somewhere around 25,000 users.

What we did

Infer the signal from the profile graph, retrieve approximately, then re-rank exactly — which is what keeps the whole thing inside 15ms.

The calls that mattered

  • Approximate retrieval, exact re-ranking

    An HNSW index over pgvector for sub-10ms nearest-neighbour retrieval, then a re-ranking pass applying mutual connections, past interactions and verification status. Precision where it changes the answer, approximation where it does not.

  • Negative feedback as a first-class signal

    Explicit 'not relevant' signals fed back into ranking. This moved the numbers more than any change to the model did.

What changed

industry average acceptance rate
3xindustry average acceptance rate
end-to-end recommendation latency
15msend-to-end recommendation latency
match relevance
4.3/5match relevance
30-day retention
62%30-day retention
  • 15ms — 12ms retrieval, 3ms re-ranking
  • 4.3/5 — user-reported
  • 62% — at 25K+ monthly actives

Three times the industry-average acceptance rate, 4.3/5 self-reported relevance, and 62% thirty-day retention across 25,000+ monthly actives.

Services used

  • Vector-based matching engine on pgvector with HNSW
  • Business-rule re-ranking layer
  • Feedback instrumentation feeding ranking weights

What we would do differently

Every project has one of these. Publishing it is the point — a case study with no regrets in it is marketing, not evidence.

The biggest improvement did not come from a better model, it came from better feedback loops. We would instrument the negative signal on day one rather than adding it once the rankings looked suspicious.

Services used

More work

Ready to start?

Start a project like thisWork