Train private models on proprietary data
Automatically spin up the entire infrastructure with production-grade AI pipelines, from data ingest, training and serving via simple code. Done in hours, not months.
Build your first AI pipelines
Complex AI pipelines done by write simple config
Train Llama-70B with Axolotl
Write a simple job configuration with task_name: Axolotl which is already integrated. Add your custom parameters, specify the training dataset, and define the output location
When the job is submitted, the platform provisions an A100 instance on VM or Kubernetes and begins training. The final model is stored in the output location you define.
name = "qlora_finetuning"description = "Finetuning Llama-7B with Axolotl"job = { task_name="axolotl", profile="node_a100", mode="train" params = { config_text="config/lora8b-instruct.yml", },}input = 'gs://bucket/input'output = 'gs://bucket.output'Generate dataset with GPT-OSS 20B
Specify the location of the notebook that performs data augmentation using GPT-OSS-20B. Include the required models in the startup configuration and set the data sources and output destination.
Once the job is submitted, the platform provisions GPT-OSS-20B automatically and executes the notebook as soon as the environment is ready.
name='generate_qa_llm'description='Augment data with GPT OSS 20B'cmd = 'python augmented.ipynb'startup = [{ name="sglang", model = "openai/gpt-oss-20b"}]input = 'gs://bucket/input'output = 'gs://bucket.output'Run any LLM model
Define the job using the pre-integrated VLLM runtime. Specify the models to be served and the required compute resources.
When the job is submitted, the platform provisions new GPU infrastructure automatically and launches the models as soon as the environment is ready.
name = "host_llm_inference"description = "Serviing custom Llama-70B"job = { task_name="vllm", params = { model="llama-70b-finetuned", profile="medium_gpu" }}Convert PDF into Markdown
Define the job to execute Python code for converting PDF files into markdown inside the Docling container. Configure the required compute resources.
After the job is submitted, the platform deploys the Docling container and runs the Python workflow automatically on the specified infrastructure.
name = 'extract_pdf'description = 'Extract PDF to Markdown with Docling'image = 'docling:latest'cmd = "chmod +x run.sh && ./run.sh"input = 'gs://bucket/input'output = 'gs://bucket/output'compute = 'single_a100'Control your pipelines via UI
Provide better view on your pipelines. Support any parallel, chaining and dependencies operations.
Run AI pipelines on your infra using DAX
Build infra automatically. Scale from VM up to Kubernetes clusters.
Your team workstation
Spin a workstation for private AI model development inside organization private networks. Secured via IAP provides protection to internal data while working from public internet.
Features
- Real-time editor online collaboration
- Pipelines building with high compute
- Produce test environment
CLI Command
dax project deployConnect via IAP
gcloud compute ssh deploy --tunnel-through-iap
Infra config as YAML
Designed for repeatable deployment across teams and environments. Translating complex compute topologies into YAML-based components. Simplify complex infrastructure for LLM and data science pipelines with consistency and precision.
Features
- VM and Clusters support
- Spot / Preemptible options for cost savings
- Overrides configuration for more advanced usage
YAML configuration
gcp_vm_g2_16: machineType: g2-standard-16 gpu: 1 osImage: projects/cos-cloud/global/images/family/cos-121-lts preemptible: "true" provisioningModel: SPOT imageSize: 50 bootSize: 30 alternativeZones: - us-east1-b - us-central1-b
Scale with K8s clusters
Fully compatible with existing Kubernetes environments or deployable on demand through DAX. Support for Ray and native Kubernetes jobs provides flexibility for a wide range of workloads. Integrated gang scheduling ensures efficient GPU allocation for high-intensity AI tasks. Operational across GKE, on-premises deployments, and any standard Kubernetes cluster.
Features
- Cloud and On-premise clusters integration.
- Jobs via Ray, AppWrapper and Kubernetes.
- Gang-scheduling for GPU compute.
- More advanced features.
YAML configuration
apiVersion: kueue.x-k8s.io/v1beta1kind: ClusterQueuemetadata: name: "cluster-queue"spec: namespaceSelector: {} # match all namespaces resourceGroups: - coveredResources: [ "cpu", "memory", "ephemeral-storage" ] flavors: - name: "default-flavor" resources: - name: "cpu" nominalQuota: 10000 # Infinite quota. - name: "memory" nominalQuota: 10000Gi # Infinite quota. - name: "ephemeral-storage" nominalQuota: 10000Gi # Infinite quota.
Favourites Features
- Container-basedRun any codes, tools and dataset inside docker image. Reduce drift issues between development and production environment.
- Cloud or Self-hostedUse DAX as control plane or runs on your own self hosted infra. Support VM-based and Kubernetes clusters for CPU + GPU.
- Enterprise SecurityBoth VM and Clusters on private network. Zero-trust policy with IAP integration. Hardened with Google Container Optimized OS technology.
- Mobile-friendlyRemote first-class. Work anywhere with laptop and pulling heavy large data task or models training just with mobile internet.
- Optimized for speedInfra provisioning in minutes, with auto-scaling, scheduler and disk caching works with air-gapped environment.
- CLI + API SupportAll complex operation done with simple command-line (CLI). Logs can be accessed via REST API logs.
FAQ
- A pipeline transforms input data into meaningful results — such as sharding large files, fine-tuning models, or running LLM evaluations. Modern LLM development relies on multiple interconnected pipelines: NeMo Curator for data preparation, Docling for extraction, Axolotl for training, and vLLM or SGLang for inference. Achieving seamless integration across these diverse tools and stages remains one of the big challenges in LLM development.
- Designing pipelines is one challenge; running them efficiently at scale is another. Even simple tasks like data augmentation with models such as Llama-70B demand significant compute — multi-core CPUs, high memory, multiple GPUs, and hundreds of gigabytes of storage. Running pipelines on distributed compute improves scalability but adds complexity, with challenges like network latency, GPU contention, driver mismatches, and unstable runtimes.
- Dagploy simplifies the entire lifecycle of AI pipelines — from creation to execution. It provides a unified system for defining workflows, provisioning compute across VMs or Kubernetes, and orchestrating GPU resources automatically. Dagploy eliminates manual setup and keeps your private LLM workloads reproducible, scalable, and secure. Its come with two offers: Daploy Pipelines and Dagploy Executor (DAX).
- A tool-agnostic pipeline builder that lets you bring any code, framework, or Docker image with custom commands and compute settings. Supports both sequential and parallel execution for flexible workflow design. Compatible with all languages (Python, Rust, etc.) and frameworks (Pandas, PyTorch, and more). Comes with a web interface to visualize, monitor, and manage pipelines in real time.
- A cloud orchestrator engine that ensures every pipeline runs reliably and efficiently. Supports both VM and Kubernetes-based computation, with optimizations like disk caching to reduce network overhead. Pipelines run under strict security standards — private network, IAP protection, and Container-Optimized OS. Configuration is defined through YAML or accessible via REST API and CLI, with auto-scaling and spot mode for cost efficiency.
- Yes. DAP and DAX provide scalable AI infrastructure for businesses of any size to train private LLMs with minimal setup, while ensuring all data remains fully secure within your own system.
- DAP is open-source, so you can use it right away without pay anything, while DAX is subscription-basis.
- Yes! There's plenty of information in the documentation.