Senior ML Backend Engineer - Search & Recommendations

Salla

About The Position

Salla is building an internally owned discovery stack for its marketplace and merchant stores, including hybrid lexical and vector search, Arabic-first query understanding, and learned ranking shared between search and recommendations. The Senior ML Backend Engineer owns the platform that makes relevance and ranking models fast, fresh, reliable, and ready to ship. Data scientists own the relevance and ranking models; this engineer owns the systems that serve them. The role sits on the ML Platform team, working across search and recommendation serving and day to day with search data scientists and the ranking lead. The work is approximately half online serving, one third streaming data, and the remainder integration into product surfaces. The engineer builds and operates search and recommendation serving, streaming pipelines that keep indexes and features fresh, experiment infrastructure, and integrations with product backends and clients. The role also includes monitoring, alerting, runbooks, and sharing on-call responsibility for systems shipped.

Requirements

At least 5 years building backend or platform systems in production.
At least 2 years operating a production search engine such as Elasticsearch, OpenSearch, Solr, Typesense, Vespa, or similar.
Experience with search index lifecycle, analyzers, large-scale reindexing, and debugging relevance issues with data scientists.
Experience building low-latency services under real traffic and explaining p95 latency.
Experience with streaming ingestion using Kafka, Debezium, or similar change-data-capture tools, including idempotent backfills.
Strong Python skills.
Comfort shipping in TypeScript.
Experience shipping changes through A/B tests and staged rollouts.
Daily use of agentic coding tools such as Claude Code.

Responsibilities

Build and operate query and recommendation serving, including retrieval, score fusion, filtering, and caching.
Meet tight latency budgets, including double-digit-millisecond p95 for search-as-you-type.
Design graceful degradation so serving remains reliable when a model or vector path is slow or unavailable.
Build experimentation infrastructure, including shadow traffic, A/B assignment, guardrail metrics, and rollback.
Build change-data-capture pipelines from source systems into search indexes and feature stores, with measured freshness.
Keep embeddings fresh for new and updated items.
Run zero-downtime reindexing of multi-million-document collections.
Build feature logging that converts production traffic into training data for ranking models.
Integrate changes into product teams' backends and clients, including new endpoints, cutovers, and experiment flags.
Coordinate with product engineering teams on reviews, release windows, and production ownership.
Provide monitoring, alerting, runbooks, and a share of on-call support for systems shipped.