We produce high-quality training data for agentic AI, embodied robotics, and foundation models — from real environments, with human expertise, at production scale.
Egocentric video demonstrations, UMI gripper data, teleoperation recordings, hand pose + SLAM trajectories. Real humans in real environments — hotels, factories, kitchens, warehouses.
Complete observe-think-act decision traces from AI agents solving real tasks in Docker-isolated terminals. Multi-step, auto-verified, SFT-ready.
High-fidelity gameplay recordings with frame-aligned inputs, cinematographic footage, and dynamic video corpora. Fuel for video generation and world simulation models.
Expert-designed evaluation suites: agent skill testing, GPU kernel benchmarks, multimodal science exams, structured image QA, engineering drawing analysis.
Structured prompt → code → rendered output triples with rubric evaluation. SVG, CSS animations, 3D scenes, interactive apps. Multi-dimensional quality scoring.
Math/science problems with CoT solutions, K12 across 5+ languages, financial agent data, medical records, traditional medicine databases, SFT corpora in Southeast Asian languages.
We work with your ML team to define task specifications, annotation schemas, quality rubrics, and acceptance criteria — aligned to your model architecture and training loop.
Domain experts and trained operators collect data in real environments — factory floors, hotel rooms, kitchens, Docker terminals — not synthetic simulations.
Automated QA (pytest, rendering checks, format validators) plus human expert review. Only data that passes both layers ships. Correction data available on request.
Your native format — Zarr, Parquet, Chat JSON, HuggingFace datasets, LeRobot, RLDS. S3-compatible API, bulk export, or custom integration.
Our operators work in actual hotels, factories, warehouses, and kitchens. The data captures real-world physics, lighting, occlusion, and human variability that simulation can't replicate.
Every dataset ships with explicit quality rubrics and automated verification. No crowd voting. No ambiguous majority labels. Deterministic, reproducible quality metrics.
Hong Kong HQ with deep operations in Mainland China and US entity (Isotope LLC). We bridge China's data execution capability with Western client standards and compliance.
We build data in your training format from day one — not collect in one format and convert later. LeRobot, GR00T, OpenAI Chat, Terminal-Bench, HuggingFace — all first-class.
Terminal-Bench standard tasks across 6 domains — algorithms, ML engineering, debugging, system admin, Git, data ops — each with Dockerfile, pytest, and auto-grading.
Egocentric demonstrations across hospitality, logistics, manufacturing, and custom SOP environments. Multi-camera, multi-annotation, production-validated pipelines.
English, Chinese, Indonesian, Vietnamese, Malay, and more. K12 curricula, SFT corpora, TTS voice libraries — each built to local education and linguistic standards.
Wearable multi-camera capture (head + dual wrist) with optional Manus haptic gloves. Real human demonstrations in real environments.
Multi-modality robotic data: handheld UMI grippers, VR/exoskeleton teleoperation, multi-view RGB-D, force/tactile sensing.
Complete observe-think-act traces in real Docker-isolated terminals. Ideal for SFT of coding and agentic models.
Every model has unique data needs. We co-design collection SOPs, annotation schemas, and QA rubrics with your team — then execute at scale in real environments. From pilot batch to production volume, we iterate until the data moves your metrics.
Define tasks, schemas, and quality rubrics with your ML team
Small-scale production for validation and iteration
Refine SOP based on your model's feedback signals
Full deployment with automated QA and continuous delivery
Zarr, Parquet, Chat JSON, HuggingFace — your native format
New Oriental Bay is a full-stack AI training data company built to serve the next generation of intelligent systems — from large language models and agentic AI to embodied robotics and reinforcement learning environments. We believe the quality ceiling of any AI system is set by its training data, and we exist to raise that ceiling.
From our Hong Kong headquarters, we coordinate data operations across Mainland China and the US, combining deep local execution capability with global delivery standards. Our teams include domain experts in robotics, software engineering, medical science, finance, and education — ensuring every dataset reflects real-world complexity, not synthetic shortcuts.
Whether you need egocentric human demonstrations captured on factory floors, agent trajectories recorded in live Docker terminals, RL environment interaction logs, or multi-language STEM problem sets verified by subject-matter experts — we design, collect, annotate, validate, and deliver. Customization is not an add-on. It's how we work.