Research Scientist Graduate (Conversational AI)- 2027 Start (PhD)

We build the next-generation unified Agent system for TikTok's global e-commerce customer service — running in 30+ languages across one of the largest e-commerce surfaces on the internet.

Our north star is a self-evolving Agent: post-training, harness, memory / context engineering, tools, and evaluation form one closed loop, and every served conversation becomes the next iteration's training / eval / retrieval / skill-induction signal. This loop is already running in production — cases are mined, root-caused, turned into constrained candidates, replayed against frozen regression sets, and shipped behind guardrails.

Two things make this team different from most "LLM application" work: - We build the agent runtime itself — Codex / Claude-Code-class — not prompts on top of a vendor API. - Evaluation and experimentation are first-class systems, not an afterthought. A self-improving loop optimizes whatever signal you give it, so the hardest and most valuable engineering here is making the judgment trustworthy — not just making the model change. As a new grad, you'll own a real end-to-end piece from day one and ship it to production.

We are looking for talented individuals to join our team. As a graduate, you will get opportunities to pursue bold ideas, tackle complex challenges, and unlock limitless growth. Successful candidates must be able to commit to an onboarding date by the end of the year. Please state your availability and graduation date clearly in your resume.

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 — 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. Agentic memory (structured note-taking), context compaction / summarization, context editing / observation masking, and just-in-time (retrieve-then-load) 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. SFT / DPO / RL to internalize rules into weights (so the prompt gets shorter, not longer), plus distillation to smaller serving models. Turn served conversations into training / eval / retrieval signals. - Tools, Skills, and MCP. Tools-as-APIs, connectors, skill / tool search for large inventories, and skill-library governance — description conflicts, trigger evals, cross-skill mis-fire matrices, and on-demand loading instead of dumping every definition into context. - Evaluation you can bet a launch on. LLM-as-judge with human-agreement calibration; statistical rigor — paired comparison, confidence intervals, repeated sampling, pass^k; held-out and time-rolling eval splits with overfitting alarms; cascaded scoring and cross-family judge panels to make evaluation affordable at scale. - The self-evolving loop. Case mining → automatic root-cause → constrained candidate generation → replay verification against frozen regression sets → canary → flywheel. Build the plumbing that makes it auditable: candidate registry with exact runtime read-back, change lineage, and an archive of rejected candidates you can sample from next round. - Online experimentation & causal readout. Shadow / canary / A-B, non-inferiority gates, traffic-split health, metric definitions that survive scrutiny, and off-policy counterfactual evaluation where live A/B isn't possible. - Safety & anti-gaming. Keep the evaluator and the release gate outside the loop that edits the system; maintain never-optimized anchor sets; monitor full execution traces rather than final answers alone; pair every quality objective with a cost-side constraint. - Own one high-leverage end-to-end surface and ship it to production across 30+ languages, measured on real