Data Scientist (ML + LLM Systems)

Intempt Technologies, LLC.; United States; On-site; USD 130,000 - 210,000; Full time

Job summary

Intempt Technologies, LLC. in the United States is seeking a production ML engineer to advance a 3-stage Think-Plan-Execute pipeline. You will optimize Flink scoring, tune embeddings, and work across a 63+ function tool-calling platform.

You should be fluent in Python, comfortable with reading Java, and able to explain model decisions to marketers. Prior experience with streaming data, Neo4j-based recommendations, and cost-per-inference awareness is valued.

Qualifications

• 3+ years building production ML systems.
• Strong Python and ability to read Java.
• Experience with streaming data (Flink, Kafka) and real-time features.
• Familiar with graph-based ML and recommendation systems.
• Experience with tool-calling agents or RAG pipelines.
• Able to explain model decisions to a marketer.

Responsibilities

• Optimize ML pipelines for cost per inference.
• Develop tool-calling agents and RAG pipelines.
• Tune Flink scoring pipeline using real-time data feeds.
• Collaborate across ML, data engineering, and product teams.
• Explain model decisions to non-technical stakeholders.

Skills

Python

Explain model decisions

Cost-per-inference

RAG pipelines

Tool-calling agents

Production ML

Tools

Flink

Kafka

Neo4j

Java

JVM

Job description

Intempt runs 9 AI agents that make real decisions for marketers every day. Not demos. Production. Real customer data.

Core Platform. Remote. Full-time. 3+ years.

Intempt runs 9 AI agents that make real decisions for marketers every day - segmenting audiences, scoring leads, predicting churn, optimizing journeys, generating content. Not demos. Production. Real customer data. Real consequences.

The brain behind Intempt's agents is a 3-stage LLM pipeline (Think, Plan, Execute) with 63+ tool-calling functions, streaming responses, and a plan-gate that asks users before side effects. Underneath that lives classical ML - Flink streaming pipelines for RFM scoring, likelihood regression, and next-best-action models running on 350M+ events/month. Plus a 17-algorithm recommendation engine backed by Neo4j graph algorithms and CLIP/SentenceTransformer embeddings.

You'll work across all of it. One day you're optimizing the Opus-to-Haiku routing logic to cut cost per turn. The next you're tuning the Flink scoring pipeline that feeds signals into the agents. If you've only fine-tuned models on clean HuggingFace datasets or only done prompt engineering, this will stretch you. We need someone who thinks in both worlds.

• Classical ML models: RFM/FA scoring (Jenks natural breaks), likelihood regression, churn prediction, next-best-action - all on Flink
• The recommendation engine - 17 algorithms spanning cart-based, popularity, affinity, and similarity approaches, running on Neo4j
• Embedding pipelines - CLIP (512-dim) for image/text similarity, SentenceTransformer for knowledge and snippet vectorization
• Eval frameworks for non-deterministic agent outputs - you define what 'good' looks like when the answer changes every time
• A/B testing and causal inference to prove models move revenue, not just improve AUC
• Model cost tracking - every inference has a dollar cost. You own the margin.
• 3+ years building production ML systems - deployed models serving real users, not research papers
• Hands-on with LLM application development - you've built tool-calling agents, RAG pipelines, or multi-model orchestration
• Strong in Python, comfortable reading Java - platform services are JVM, ML layer is Python
• Experienced with streaming data and real-time feature engineering (Flink, Kafka, or similar)
• Familiar with graph-based ML and recommendation systems - bonus if you've worked with Neo4j
• Someone who obsesses over cost-per-inference as much as accuracy - running Opus on every turn isn't an option
• Able to explain a model's decision to a marketer who's never heard of a confusion matrix
• You've only done prompt engineering or only done classical ML - we need both in one person
• You think shipping a Jupyter notebook to production is acceptable