01

Matching.

Recommendation systems for human compatibility.

42ms
p99 ranking latency
1.4M
candidates per query
What we build

Recommendation systems for human compatibility. Productionized.

Multi-signal matching engines that combine collaborative filtering, embedding similarity, and explicit preference logic. We train, evaluate, and ship the entire pipeline, from feature extraction to the reranker that decides what shows up at the top of the feed.

Common deployments

Our dating brandsCommunity discoveryMember recommendations

What we ship

  • Production ranking service with A/B harness wired in.
  • Reviewer dashboards for ML drift and quality monitoring.
  • Feature store with documented schema and SLAs.
  • Runbook for retraining cadence and rollback.
System overview

Five components, one cohesive runtime.

We build each layer to operate independently, then compose them into the brands we run for our members.

01layer 1 of 5

Feature pipeline

Streaming + batch features, served from a low-latency store.

02layer 2 of 5

Embedding service

Vector encoders for profile, behavior, and content signals.

03layer 3 of 5

Two-tower retrieval

Approximate nearest neighbor search at sub-50ms p99.

04layer 4 of 5

Reranker

Gradient-boosted reranker tuned to business-defined success metrics.

05layer 5 of 5

Feedback loop

Implicit and explicit signal ingestion with A/B harness.

Tech we reach for

We pick boring, proven tools where boring works. Where the problem demands more, we build deliberately on top of these primitives.

pgvectorFAISSRedisKafkaPythonGo
Outcomes

What it looks like in production.

+38%

Day-7 retention lift (typical)

<50ms

End-to-end recommendation latency

94%

Recommendation acceptance rate

Get in touch

Building the systems that bring people together.

Questions about our brands, press, or partnerships? Send a note and we'll reply within two business days.