Software Engineer, Full-Stack

About the job

Responsibilities

• Design context, tools, skills, and orchestration for multi-step agents, including work across documents, slides, images, audio, and video.
• Build self-improvement loops with explicit permissions, evaluation gates, and rollback mechanisms.
• Curate trajectories for SFT, optimize preferences, and run RL on real agent tasks, investigating methods like DPO, GRPO, or DAPO.
• Compare process and outcome supervision, shape rewards, and distill useful frontier behavior into smaller models.
• Measure whether gains transfer beyond the training environment.
• Turn enterprise workflows into reproducible training and evaluation environments, including fixture tenants and simulated users.
• Find shortcuts an agent can exploit before a training run optimizes for them.
• Mine production agent-steps for failures and build curated corpora with useful synthetic augmentation.
• Calibrate judges against human labels and construct behavior-level benchmarks from real workflows.
• Quantify data quality, performance uplift, and reliability across stochastic runs.
• Connect enterprise information with user- and tenant-level learnings.
• Separate failures of retrieval from failures to use retrieved context.
• Test what to retain, update, and retrieve to improve future decisions.
• Build and evaluate routing, ensembles, caching, and small-model specialization.
• Measure downstream task success alongside latency and cost.
• Define problems and baselines, build data or environments for testing, and run experiments.
• Own serving and integration of developed systems.
• Follow systems through deployment, monitoring, and failure analysis.
• Collaborate with researchers and engineers across different locations.

Requirements

• Substantial work in at least one of agent/tool-use systems, post-training, reward modeling or RL environments, retrieval and memory, or evaluation design.
• Demonstrated ability to take a system, model, or research project from an ambiguous problem to a working result.
• Ability to size experiments, choose meaningful baselines and held-out tests, and account for variation.
• Capacity to distinguish real improvement from bias, data leakage, or benchmark shortcuts.
• Ability to debug across model and system boundaries, isolate failures, and produce maintainable production code.
• Commitment to honest measurement and retiring approaches with clean negative results.
• A Master's or PhD in a relevant field, or equivalent work or research experience.
• Demonstrated depth through papers, open-source contributions, trained models, or well-documented experiments.