Recruiting · Production

Malt super-powers freelancer recommendation with retriever-ranker architecture and Qdrant vector database

The problem

Malt's original monolithic matching model had response times of up to one minute and was inflexible, making it difficult to adapt for future large language models or scale to real-time recommendation needs.

Workflow diagram · grounded in source
1
Project posted, vector encoded
trigger
“When a new project is posted, its details are encoded into a vector in real-time.”
2
Retriever narrows freelancer pool
ai_action
“The Retriever is built for recall, ensuring that we do not miss out on potential matches, even if, at this step, we are not yet state of the art in terms of quality”
3
Ranker scores and reorders subset
ai_action
“the ranker, will rank the 1,000 profiles and place at the top of the list the freelancers that perfectly match the project (Java version, industry etc…)”
4
Near-real-time recommendations delivered
output
“enabling us to recommend freelancers in near real-time, significantly enhancing the user experience”
Reported outcome

After deploying the retriever-ranker architecture backed by Qdrant, p95 latency fell from tens of seconds (sometimes over a minute) to 3 seconds at most, and an AB test confirmed an increase in project conversion without sacrificing recommendation quality.

Reported metrics
P95 response latency3 seconds at most
Project conversionincrease in the conversion of projects in our AB test
Reported stack
QdrantKubernetesGrafanaPrometheusDocker
Source
https://blog.malt.engineering/super-powering-our-freelancer-recommendation-system-using-a-vector-database-add643fcfd23
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

After deploying the retriever-ranker architecture backed by Qdrant, p95 latency fell from tens of seconds (sometimes over a minute) to 3 seconds at most, and an AB test confirmed an increase in project conversion with…

What tools did this team use?

Qdrant, Kubernetes, Grafana, Prometheus, Docker.

What results were reported?

P95 response latency: 3 seconds at most; Project conversion: increase in the conversion of projects in our AB test (source-reported, not independently verified).

How is this recruiting AI workflow structured?

Project posted, vector encoded → Retriever narrows freelancer pool → Ranker scores and reorders subset → Near-real-time recommendations delivered.