Conversational AI ML Engineer Intern (PhD)

TikTok; Seattle (WA); On-site; USD 42,000 - 54,000; Part time

Benefits offered by this job

Health insurance

Wellbeing benefits

Paid holidays

Job summary

TikTok is seeking a Machine Learning Engineer Intern (Conversational AI) for a 2027 start in Seattle. You will help build agent runtimes, memory, and context handling for multi-turn conversations, and contribute to production-safe prompts and guardrails.

Strong Python plus exposure to C++/Go/Rust/Java is expected. Responsibilities span post-training data flywheels, tool search governance, and evaluation pipelines using LLM-as-judge concepts.

Qualifications

• Pursuing a PhD in a technical field (CS/Engineering/OR).
• Strong Python with one of C++, Go, Rust or Java.
• Solid ML/DL/NLP fundamentals with hands-on LLM/agent experience.
• Ability to interpret research into production-ready code.
• Reading papers/blogs and turning insights into working implementations.

Responsibilities

• Architect agent runtime: manage skills, tools, context, and guardrails; implement safe fallbacks.
• Design memory and context strategies for long multi-turn agents.
• Lead post-training data workflows: SFT/DPO/RLHF/RLAIF through distillation.
• Develop tools, MCP, and tool search with governance for scalable context.
• Evaluate via LLM-as-judge with rigorous calibration and test splits.
• Build auditable loops: candidate registry, runtime traceability, and change lineage.
• Execute online experimentation: shadow/canary, A-B tests, and counterfactual eval.

Skills

Python

C++

Go

Rust

Java

Education

PhD candidate in Computer Science/Engineering/Operations Research

Job description

TikTok is seeking a Machine Learning Engineer Intern (Conversational AI) for a 2027 start in Seattle. You will help build agent runtimes, memory, and context handling for multi-turn conversations, and contribute to production-safe prompts and guardrails.

Strong Python plus exposure to C++/Go/Rust/Java is expected. Responsibilities span post-training data flywheels, tool search governance, and evaluation pipelines using LLM-as-judge concepts.