Senior AI/ML Engineer

We're hiring a Senior AI/ML Engineer for the Discovery Pod to help build and run the ranking, recommendation, and search systems buyers rely on to find the right stone or piece in seconds. This is real information retrieval and applied ML work on production infrastructure — search systems, embedding-based retrieval, learning-to-rank, LLM-powered contextual and conversational search, image embeddings for photo-based search — not configuring a hosted search widget or calling a model API and moving on. You'll own models end to end, from features and training through evaluation, deployment, and monitoring, and you'll be expected to understand how these probabilistic systems behave and fail, because that understanding is the job, not an adjacent skill.

TECH STACK

- Search: OpenSearch, fed via CDC from PostgreSQL plus Spark/EMR/Glue ingestion jobs. No vector database anywhere in production — embedding-based retrieval, the ranking model, and image-embedding search are genuinely new builds for whoever takes the Discovery role.

- Core DB / warehouse: PostgreSQL (OLTP) → Snowflake (dbt-modeled, medallion raw/curated/presentation layers)

- Backend: TypeScript/Node.js, GraphQL API gateway, Nx monorepo + PNPM

- Auth: Keycloak · SSO: Okta

- CI/CD: Jenkins (“Nivoda Build”) + GitOps for deploys

- Cloud: AWS (EMR, Glue, DMS, Kubernetes-based deploys)

- Observability: Grafana for logs/traces

WHAT WOULD YOU BE DOING?

Core Search Engine

- Build and operate real search infrastructure — Elasticsearch/OpenSearch, Vespa, or vector-search databases like pgvector, Pinecone, or Weaviate — to power ranking and multi-modal search across Nivoda’s marketplace

- Design multi-tenant, configurable ranking and recommendation systems that serve both Nivoda’s own marketplace and many individually branded jeweller storefronts through Feeds, without forking the codebase

Multi-Modal AI

- Build and improve Contextual Search and our planned conversational assistant, using prompt design and retrieval-augmented approaches, with zero-result and tradeoff scenarios handled gracefully rather than met with a confident wrong answer

- Build image-to-attribute and image-to-product-match pipelines using image embeddings for our image-upload search work

- Build the product around the reality that probabilistic systems can be confidently wrong — designing recovery paths and confirmation steps before high-stakes actions, like generating a manufacturing quote from an interpreted image