Senior Machine Learning Engineer

What You’ll Do Tekmetric sits on a decade of granular operational data from thousands of automotive repair shops — millions of operational records spanning customer history, service patterns, and technician activity. We are building the ML layer on top of it: retrieval systems, recommendation models, embedding infrastructure, and the LLM integration layer that makes AI features feel native to the product workflow. This is a systems role for someone who builds production AI seriously.

• Design and build RAG pipelines — chunk store schema, Redshift-sourced population, and the signal-based retrieval layer that assembles context for LLM-powered suggestions in real time.

• Own the semantic search infrastructure over operational data: embedding pipeline, chunking strategy, hybrid BM25+vector retrieval, and the reranking layer that feeds LLM tool calls (PGvector).

• Build a recommendation model end-to-end — training pipeline (feature extraction, model development, evaluation) and the SageMaker inference path that surfaces relevant suggestions at the right moment in the user workflow.

• Design tool definitions, prompt assembly, and context selection for Bedrock-backed AI features, working directly with application engineers who wire these into the product.

• Compute and serve ML features that recommendation models and customer analytics depend on: entity-level historical aggregates, account behavioral profiles, and engagement history.

• Build and maintain batch and incremental embedding pipelines via Bedrock (Titan, Cohere) with CDC-based freshness, keeping the vector store current with the operational database.

• Build evaluation frameworks — offline retrieval metrics, model quality tracking, and A/B infrastructure needed to measure whether AI features are driving business outcomes.

• Collaborate with data engineers to integrate ML models into data pipelines and surface results through search and analytics APIs.

What You’ll Bring

• 4+ years in ML engineering with clear experience taking a model or ML feature from experimentation into production — not just research or notebooks.

• Retrieval systems depth — embeddings, vector similarity, and ANN search at the mechanism level; ability to reason about BM25 vs. dense retrieval, chunking tradeoffs, and when reranking adds value (PGvector, FAISS, Qdrant, Weaviate, Pinecone).

• LLM application patterns — hands-on experience with RAG, tool use/function calling, prompt engineering, and context assembly; debugging retrieval pipelines where model output was wrong because the context was wrong (Bedrock, OpenAI, Anthropic SDK, LangChain).

• Strong Python and ML tooling — production-quality training and inference code, not just exploratory notebooks (PyTorch, scikit-learn, Hugging Face, XGBoost).

• Data engineering fundamentals — writing the SQL or Spark job that extracts training data without handing it off; understanding point-in-time correctness and feature leakage.

• ML infrastructure on AWS — deploying and operating models in a real cloud environment (SageMaker, Bedrock, S3, Lambda).

• Workflow orchestration — building ML pipelines that run on a schedule, handle failure gracefully, and do not require manual intervention to stay healthy (Airflow, Prefect, or Kubeflow).

Nice to Have

• Fine-tuning or PEFT — experience adapting embedding or generative models to a domain-specific corpus.

• Graph databases — Apache AGE, Neo4j, or similar; understanding how graph traversal complements vector similarity for multi-hop relationship queries.

• Streaming and CDC — Kinesis, Kafka, or Debezium; keeping a vector store or feature store current with a live operational database.

• Recommendation systems — collaborative filtering, two-tower models, or contextual bandits in a production setting.

• Domain experience in operational data — automotive, field service, logistics, healthcare; any domain where records are sparse, the vocabulary is specialized, and keyword search reliably fails.