Research Scientist Graduate (Conversational AI)

About the Role

We are building the next-generation unified Agent system for TikTok's global e-Commerce customer service. Our goal is a self-evolving Agent where post-training, harness, memory/context engineering, tools, and evaluation form a closed loop, with each served conversation becoming a signal for the next iteration. This loop is already in production, mining cases, performing root-cause analysis, generating constrained candidates, and shipping behind guardrails.

This team differs from typical "LLM application" work by:

• Building the agent runtime itself (e.g., Codex / Claude-Code-class), rather than relying on prompts on top of vendor APIs.
• Treating evaluation and experimentation as first-class systems.

As a graduate, you will have the opportunity to pursue bold ideas, tackle complex challenges, and achieve limitless growth.

Key Responsibilities

• Agent runtime (harness / agent loop): Orchestrate skills, tools, and context; implement loop control & intervention, progressive disclosure, and behavior-level guardrails. Build the production safety layer including pre-flight budgets and timeout truncation, serve-time gates, shadow / swap-in answer delivery, and safe fallback paths.
• Context & memory for long multi-turn agents: Implement agentic memory (structured note-taking), context compaction / summarization, context editing / observation masking, and just-in-time retrieval. Treat context as an evolving, itemized playbook with structured diffs and a deterministic curator, rather than an ever-growing prompt.
• Post-training & the data flywheel: Utilize SFT / DPO / RL to internalize rules into weights, and distillation to smaller serving models. Transform served conversations into training / eval / retrieval signals.
• Tools, Skills, and MCP: Develop Tools-as-APIs, connectors, skill / tool search for large inventories, and skill-library governance including description conflicts, trigger evals, cross-skill mis-fire matrices, and on-demand loading.
• Evaluation: Implement LLM-as-judge with human-agreement calibration; use statistical rigor including paired comparison, confidence intervals, repeated sampling, and pass^k; manage held-out and time-rolling eval splits with overfitting alarms; utilize cascaded scoring and cross-family judge panels for scalable evaluation.
• The self-evolving loop: Build the auditable plumbing for case mining → automatic root-cause → constrained candidate generation → replay verification against frozen regression sets → canary → flywheel, including a candidate registry with exact runtime read-back, change lineage, and an archive of rejected candidates.
• Online experimentation & causal readout: Implement shadow / canary / A-B testing, non-inferiority gates, traffic-split health checks, robust metric definitions, and off-policy counterfactual evaluation where live A/B isn't feasible.

Requirements & Tech Stack

• Strong Python
• Proficiency in one of C++ / Go / Rust / Java
• Solid ML / DL / NLP fundamentals
• Genuine hands-on experience with LLMs or agents (coursework, research, internship, competition, open-source, or a serious side project)
• Basic statistical literacy (confidence intervals, p-value interpretation, distinguishing correlation from causation)
• Ability to translate research papers or engineering blogs into working code

Preferred Qualifications

• Experience building agent runtimes, not just calling APIs (e.g., own agent loop/harness, memory/context system, RAG or tool-use agent, or a fine-tuned/post-trained model).
• Hands-on experience in any of the following:
• Post-training: SFT / DPO / RLHF / RLAIF / RLVR, reward modeling, reward hacking defense.
• Agent systems: Harness, context engineering, MCP / Skills, sub-agents, tool search.
• Evaluation & experimentation: LLM-as-judge and judge calibration, pass^k, regression suites, A/B and non-inferiority testing, off-policy evaluation.
• Self-improving / evolutionary systems: Evolutionary program search, candidate archives and parent sampling, automatic prompt/context optimization, multi-objective (Pareto) selection and credit assignment.
• Inference & serving: vLLM / TensorRT-LLM, MoE, KV / prompt caching, cost engineering.
• Multilingual NLP
• Publications (for PhD), strong competition results (ACM-ICPC / Kaggle / ML competitions), or notable open-source contributions.
• Demonstrated preference for rigorous analysis over premature conclusions (e.g., have stopped an experiment due to lack of statistical significance).
• e-Commerce or multilingual experience is a plus, not required.

Qualifications

• Individuals who are completing or have recently completed a PhD degree in Computer Science, Electrical Engineering, Mathematics, Statistics or a related discipline.