Data Science Workflows
On-premise GPU hardware is more cost-effective than cloud for sustained training workloads — but only if it's actually accessible. The Agentic API gives ML engineers scoped access to GPU machines and data directories without exposed ports, without a DevOps ticket, and without cloud egress fees.
The Problem
Getting Safe, Scoped Access to Local GPU Hardware Takes Too Long
Data science on local or on-premise GPU hardware has a persistent access problem. Exposing a Jupyter port is a security risk. Granting full SSH access is too broad — a data scientist does not need to touch system files to run a training job. Routing everything through cloud infrastructure is expensive and slow when your datasets are hundreds of gigabytes. And getting a DevOps team to provision correct access for each experiment takes time that kills iteration speed.
Jupyter Ports Are a Security Risk
Exposing port 8888 to the internet is a known attack surface. Notebook state drift and lack of version control make reproducibility difficult in addition to the security exposure.
Cloud Egress Is Expensive at Dataset Scale
Moving hundreds-of-gigabyte training datasets to cloud compute adds significant egress costs and round-trip latency. Your own GPU hardware is more cost-effective — if you can reach it cleanly.
How the Agentic API Solves This
Scoped GPU Access From Any Environment.
The Agentic API gives data scientists safe, scoped access to GPU machines and data directories — without exposed ports, without a DevOps ticket, and without cloud egress fees. A project scoped to relevant training directories and specific script commands gives an ML engineer exactly what they need: push datasets, trigger training runs, pull results. Nothing else. Callable from a laptop, a cloud CI job, or a Python script.
Step 1 — Define the Data Science Scope
Create a project scoped to training, dataset, config, and results directories. Allowlist specific training scripts and evaluation commands. Retrieve Project Key and Secret.
Step 2 — Stage, Trigger, Harvest
POST preprocessed datasets to the GPU machine's data directory via the file API. POST the training command to start the run. After completion, retrieve checkpoint files and metrics back to your orchestrator.
Step 3 — Iterate Without Interactive Sessions
Replace ad-hoc Jupyter notebooks with version-controlled scripts triggered via API. Exit codes provide pass/fail signal; the API call log serves as the experiment audit trail. No exposed port 8888.
Use in Practice
From Hyperparameter Sweeps to Result Harvesting.
Remote GPU Training Trigger
POST a training command to a dedicated GPU machine from any Python script or CI job. Receive training logs in the response body. Exit code 0 confirms successful completion.
Dataset Staging Without SCP/SFTP
POST preprocessed datasets directly to the GPU machine's data directory via the file API. No SCP client, no SFTP configuration, no SSH tunnel required.
Hyperparameter Sweep Orchestration
Loop over parameter configurations and fire one API call per config in parallel. Collect stdout (validation loss) and exit codes to rank configurations without any interactive session.
Result Harvesting
After training, retrieve checkpoint files, evaluation metrics JSON, and generated plots back to your orchestrator via the file API. Large files can be streamed.
Recommended Directory Structure
Scope Your Project to These Paths.
| Directory | Access Mode | Purpose |
|---|---|---|
/home/awaberry/datasets/ | Write from pipeline | Preprocessed training data |
/home/awaberry/training/ | Read-only | Training scripts (prevent tampering) |
/home/awaberry/configs/ | Write from orchestrator | YAML/JSON experiment configs |
/home/awaberry/checkpoints/ | Read back to orchestrator | Model checkpoints |
/home/awaberry/results/ | Read back to orchestrator | Metrics, plots, reports |
Use Cases
More Use Cases
← Previous
Security & Compliance
Grant auditors and incident responders time-limited, purpose-scoped access with a project key. Access ends the moment you delete the project — no user account cleanup required.
Next →
Home Lab & Personal Infrastructure
Access your Raspberry Pi, NAS, or home server from anywhere — no DDNS, no port forwarding, no VPN configuration. Install the agent and your hardware becomes a clean authenticated endpoint.
Products
Explore What awaBerry Offers
Products
awaBerry Anywhere
awaBerry Anywhere is a zero-trust remote access platform that gets any device — cloud server, laptop, or SoC hardware — securely accessible from anywhere in minutes. No VPNs, no open inbound ports, no complex configuration or additional remote connection software. Works on any MAC - yes even an old Apple macbook from 2012. Works on any Ubuntu / Debian / Redhad based LINUX. Works on any Windows which supports the Windows Subsystem for Linux (WSL).
Connect Devices
Flexible onboarding for any hardware — in any environment.
#Linux
#macOS
#Server
#Docker containers
#SoC devices
#True Headless Device Installation
#Windows
#Bespoke OS installation
#Shared device access
Read more →
Secure Remote Access
Full control and activation via the awaBerry web dashboard.
#Web SSH
#Anywhere Shell
#Zero-Trust Port Forwarding
#No VPN
#Remote Desktop (VNC/RDP)
#Securely Connect To Any Port
#NoOpenFirewallPorts
#Web-based file transfer
#Zero-Trust
#Web-to-local
Read more →
Products
awaBerry Device Automation
awaBerry Automation is the combination of two tightly integrated products that together form a complete, AI-native automation platform.
#Local Execution
#Gemini CLI
#Gemini CLI GUI
#Zero Token Cost
#Agentic API
#AI-Powered Automation
#Encrypted Tunnels
#Device Fleet
#Plain English
#Scheduled Tasks
#Device-access-as-a-Service
Smart Automation Framework
Uses the Google Gemini CLI to translate plain-English instructions into executable scripts — run on your local devices on a schedule. AI tokens are spent exactly once to generate the logic; every subsequent execution costs nothing.
Read more →The Agentic API — Secure, programmatic device access
Secure, zero-trust device access as-a-service. Exposes your registered devices to programmatic access via encrypted tunnels — grant AI agents, scripts, or collaborators precisely scoped access.
Read more →The Power of Combination — Intelligence meets reach
Fundamentally different: a complete, AI-native automation platform — across every device you own, anywhere in the world.
Read more →Plans
Start for Free. Scale as You Grow.
Business
Automate operations, manage fleets, and enable secure remote work for your entire team.
Try for Free
Private
Access your home devices, automate personal tasks, and share access with family — for free.
Freemium Plan Available
Research
Access lab hardware, automate data collection, and collaborate across institutions.
Request Free Usage
Make your GPU hardware a first-class part of your ML pipeline
Scoped access from any environment. No Jupyter port, no DevOps ticket, no egress fees.
