Ml Engineer

Role & responsibilities

About the role:

A knowledge-graph retrieval system that ingests multi-source data (documents, Jira, Salesforce and more) at large scale (100K+ records/day), extracts a structured graph, and answers questions over it. You'll own the modeling and evaluation layer that decides whether the system retrieves the right thing, resolves entities correctly, and runs affordably at scale - hands-on across LLMs, information retrieval, and graph ML.

What you'll own

• Extraction quality - improve LLM/DSPy entity, relationship, concept and proposition extraction via prompt optimization, model selection, and structured-output evaluation.
• Entity resolution at scale - embedding-based dedup with threshold tuning, blocking/partitioning, and false-merge/false-split trade-offs across sources.
• Retrieval & ranking - graph-boosted reranking (centrality, edge-weight, neighbor signals), retrieval-mode selection, and intent classification driving traversal.
• Evaluation harness - labeled eval sets and the metrics that gate every change: retrieval hit-rate, intent accuracy, reranking lift, contradiction/consistency rates.
• Cost & latency - own the cost/quality frontier: deterministic vs LLM extraction, batch embedding throughput, model tiering, caching.

What you'll do day to day

• Run offline/online experiments (A/B harnesses) to validate that a change improves answer quality.
• Build and maintain labeled datasets and CI metrics tracking model/pipeline performance over time.
• Partner with backend engineers (graph store, connectors, APIs) to take models to production.
• Tune thresholds, weights, and prompts against real query traffic and report measurable impact.

Required

• 3+ years building ML systems in production (not just notebooks/research).
• Degree in CS/ML or related field from a top-tier university.
• Strong Python; comfortable in a production codebase with tests and CI.
• Hands-on with LLMs (prompting, structured output, evaluation) and embeddings / vector search (Milvus, FAISS, pgvector).
• Solid IR / ranking grounding and evaluation methodology (labeled sets, precision/recall, A/B testing).
• Entity resolution, deduplication, or record linkage at scale.
• Excellent written and verbal communication - able to explain trade-offs to engineers and stakeholders.

Nice to have

• Graph databases (Neo4j/Cypher) and graph algorithms (centrality, community detection).
• RAG / knowledge-graph systems; agent frameworks (LangGraph, DSPy).
• International experience working with distributed / cross-geography teams.
• Data-pipeline and cost/latency optimization at high volume; MLOps (experiment tracking, versioning, monitoring).

Only Immediate Joiner