Data Panda lands your operational and warehouse data in one place and ships it to an Anyscale Ray cluster. Train models, fine-tune LLMs, run batch inference and serve online endpoints, all on the data your business already generates.
Anyscale was founded in 2019 by Robert Nishihara, Philipp Moritz and Ion Stoica, the team that built Ray at UC Berkeley's RISELab between 2016 and 2017. Ion Stoica also co-founded Databricks and chairs Anyscale today; Philipp Moritz is CTO and Robert Nishihara is the third co-founder. Keerti Melkote took over as CEO in 2024.
The product is the Anyscale Platform: a managed, multi-cloud runtime for Ray that runs on AWS, GCP, Azure, CoreWeave and Nebius. It wraps the Ray libraries (Ray Data for distributed data processing, Ray Train for model training, Ray Serve for online inference, Ray Tune for hyperparameter search and RLlib for reinforcement learning) with workspaces, jobs, services, observability, governance and priority-aware GPU scheduling. OpenAI runs Ray to coordinate training of its largest models, including ChatGPT, and Anyscale's own customer list includes Coinbase, Character.ai, Canva, Notion, Runway, Grab, Recursion, TripAdvisor, TwelveLabs, Riot Games and Physical Intelligence. The open-source Ray project has over 41,000 GitHub stars and more than 500 million downloads.
GPU-hours and Anyscale spend per job, next to the model the job produced and the business workflow that uses it.
Ray Serve scores or batch predictions written straight into the systems your team works in.
Pull warehouse content into a Ray cluster, fine-tune or train, write the artifact to model storage.
Custom tools that read warehouse data and call a Ray Serve endpoint instead of a public LLM API.
A list of concrete reports, automations and AI features we have built on Anyscale data. Pick the one that matches your situation.
Which Ray jobs are spending the most, and is the model output being used?
GPU-hours and Anyscale spend per job and per workspace, joined to the model artifact each job produced and to the downstream workflow that calls the model. Surfaces the weekly fine-tune burning the bulk of the cluster budget while its model is still pinned to a stale version, next to the lighter Ray Tune sweep that produced the recommender currently running in production.
Are our Ray Serve endpoints still earning their GPUs?
Ray Serve request volume, latency and error rate per endpoint, joined to GPU usage and cost. Catches the always-on endpoint serving five calls a day on two reserved GPUs, and the busy endpoint that is autoscaling well past its budget because every internal app added a retry loop.
Is the data we are feeding the cluster fresh enough?
Lineage from the source system through the warehouse table to the dataset version a Ray Train job consumed, with timestamps at every hop. Catches the fine-tuning run that has been training on a feature table that stopped refreshing after a pipeline failure two weeks ago, which is why the model keeps regressing on last month's behaviour.
GPU-hours and Anyscale spend per team, per workspace and per business unit, joined to the workflow each model serves. The AI line on the budget moves from a single Anyscale invoice to a number sitting next to the predictions consumed in CRM, ERP and the warehouse.
Ray Serve endpoint health, latency and cost per call, joined to the internal app calling the endpoint. Lets the team retire the always-on endpoint that nobody calls and rightsize the one that every retry loop is hammering.
You keep the reporting tool you already have. We connect it to the warehouse where your Anyscale data lives.
OAuth authentication. Read-only by default. We sign a DPA and your admin keeps the keys.
Data flows into your warehouse on your schedule. Near real time or nightly, your call. You own the data.
We build the first dashboard, workflow or AI feature with you, then hand over the keys. Or we stay on for ongoing delivery.
We set up the foundation. Your team builds on top.
Best fit Teams that already have a BI analyst or data engineer and want to own the build.
We build the whole thing, end to end.
Best fit Teams without in-house BI or dev capacity. You tell us what you need and we deliver it.
You do. It lands in your warehouse, on your cloud account. We don't resell or aggregate it. If you stop working with us, the warehouse stays yours and keeps running.
Near real time for most operational systems. For heavier sources we schedule hourly or nightly. You pick based on what the reports need.
No. If you don't have one, we help you pick one and set it up as part of the first delivery. Common starting points are Snowflake, Microsoft Fabric, or a small Postgres start.
Ray is the open-source distributed compute framework: Python APIs plus the libraries Ray Data, Ray Train, Ray Serve, Ray Tune and RLlib. You can run Ray yourself on your own Kubernetes or VMs. Anyscale is the company behind Ray and runs the managed Anyscale Platform: a multi-cloud runtime for Ray with workspaces, jobs, services, observability, governance and priority-aware GPU scheduling on AWS, GCP, Azure, CoreWeave and Nebius. Same Ray code, less cluster operations work.
Anyscale runs on AWS, GCP, Azure, CoreWeave and Nebius. The same Ray code is portable across them, and the platform supports multi-cloud GPU pooling so a job can burst to whichever provider has capacity. The cluster lives in your own cloud account in customer-hosted mode, which is the typical setup for teams that need data and compute to stay in their own VPC.
Distributed model training (Ray Train), batch inference and multimodal data preparation (Ray Data), online inference (Ray Serve), hyperparameter search (Ray Tune) and reinforcement learning including RLHF post-training (RLlib and community frameworks like SkyRL and veRL). The same cluster usually carries multiple of these in production, which is why workspace-level cost and usage reporting matters more than per-job timing.
We review your Anyscale setup and the systems around it. Together we pick the first thing worth building.