Apple Technology/Software

Senior Machine Learning Engineer, Conversational AI - Services Special Projects

Cupertino, CA In person Full-time

At Apple, great ideas have a way of becoming phenomenal products, services, and customer experiences very quickly. We're building a massive, real-time conversational experience from the ground up, one that will reach users at Apple scale. It sits at the intersection of Generative AI, Information Retrieval, and Knowledge Graphs.

We're looking for a senior machine learning engineer to own the conversational intelligence layer: the agent's reasoning and tool-calling, the low-latency LLM inference behind it, and the knowledge services that keep every response accurate and grounded. You'll partner closely with teams across Apple to bring it to production.

This is a hands-on role for an engineer who has shipped conversational AI and LLM systems to production and cares about latency as much as model quality. You'll work closely with AI/ML scientists and engineers across the stack. You'll design multi-turn conversational experiences, build agent logic that plans and calls tools, and develop query understanding models that turn natural language into structured intent. You'll also make large and small models fast and cost-efficient enough to serve in real time.

Just as importantly, you'll help make sure every response is accurate and grounded. You'll build and scale knowledge graph services that resolve, enrich, and expand user requests with low latency and also connect the system to real-time data sources so its answers stay current. Finally, you'll build the evaluation that keeps all of it honest: automated and LLM-as-judge evaluation for multi-turn conversations, built into the development workflow so regressions are caught before they reach users.

Orchestrate tools and grounding: Own the agent's planning and tool-calling layer, including tool schemas, dynamic tool selection, and structured output generation. Ground every response in retrieval, knowledge graph, and real-time data.

Build conversational understanding: Design and ship conversational flows, including intent understanding, context tracking, and disambiguation, using LLM- and small-model-based query parsers that feed downstream retrieval and ranking.

Optimize LLM inference: Profile and reduce serving latency and cost through distillation, quantization, compact output formats, dynamic batching, KV-cache tuning, and semantic caching. Use inference runtimes such as vLLM, SGLang, TensorRT-LLM, and ONNX Runtime, and decide when a request needs a large model, a small model, or none at all.

Build autograded evaluation: Design and own automated evaluation for conversational and generative quality, including multi-turn rubrics, LLM-as-a-judge pipelines, judge calibration against human ratings, and error-analysis tooling. Integrate eval runs into CI.

Connect offline metrics to online outcomes: Define offline metrics that predict user-facing wins, and partner on the A/B tests that confirm them.

Ship models and services to production: Take models and services from experiment to the live request path with proper versioning, monitoring, and rollback, and own their quality, latency, and reliability after launch.

Collaborate cross-functionally: Work closely with research scientists, engineers, and product teams to turn recent research in conversational AI, retrieval, and inference into systems that hold up in production.

Master's degree in Computer Science, Machine Learning, NLP, or a related field

7+ years of experience as a software or machine learning engineer, with a significant focus on conversational AI, search, or LLM-powered systems

Strong programming skills in Go and Python, plus at least one of Java or Scala

Hands-on experience building and shipping conversational AI or dialogue systems: intent and slot understanding, multi-turn context management, clarification, and disambiguation

Experience building LLM-powered applications in production, including prompt design, tool/function calling, structured output generation, and agentic orchestration patterns

Demonstrated LLM inference-optimization experience: distillation, quantization, batching, KV-cache tuning, caching, and profiling to find real bottlenecks rather than guessing at them

Hands-on experience deploying and serving LLMs or ML models directly in the production request path, using serving frameworks such as vLLM, SGLang, Triton, or TorchServe

Experience with knowledge graphs and low-latency lookup and retrieval services

Fluency with PyTorch and the Hugging Face ecosystem (transformers, PEFT)

Familiarity with cloud environments (including AWS) and containerization (Docker, Kubernetes)

Excellent communication skills and a collaborative mindset

Ph.D. in Computer Science, Machine Learning, NLP, or a related field

Experience fine-tuning and distilling LLMs for task-specific use, such as query parsing, classification, and response generation, using techniques such as LoRA/adapters and teacher/student setups

Experience designing knowledge graph schemas, building graph compilation and ingestion pipelines, and working with graph databases such as TigerGraph or Neo4j

Deep experience with KV stores and in-memory serving for low-latency lookups (e.g., Cassandra, Redis, SSTables)

Experience calibrating LLM judges against human judgments for multi-turn conversations, including inter-annotator agreement and judge reliability tracking

Experience with personalization and memory in conversational systems, and applying privacy principles such as data minimization and on-device processing

Experience implementing safety guardrails for LLM-powered agents, including prompt-injection defenses and adversarial evaluation

Familiarity with experiment tracking and data/model versioning tools (MLflow, Weights & Biases, DVC)

Deep understanding of distributed systems and service design, and experience architecting and leading projects with cross-functional stakeholders

Published work or patents in conversational AI, information retrieval, knowledge graphs, or efficient inference

Job description provided by Apple.
View the original posting