rockie · academic applications
Before Pebble built environments for companies, it built one for its own work. Rockie is that environment: a persistent research system where experiments are planned, run, evaluated, and branched, with every job tracked, budgeted, and recorded. It has run Pebble's research program around the clock ever since, and academic groups use it for HPC research, including ML. The commercial practice on the rest of this site grew out of it.
a product of pebble ml · self-serve for academic researchers · your models & compute
HPC experiments run for hours or days. They fail, drift, and stall while no one is watching. Rockie tracks each job, reads its logs, and pulls down the artifacts it produces, then keeps the loop moving.
While one experiment runs, Rockie prepares likely successors for the results it may produce, so compute never idles waiting for someone to decide what's next.
Shared metrics and acceptance rules connect each result to the next experiment. An append-only decision journal records why every branch was taken.
Independent branches move to parallel workers without losing lineage or duplicating work. Add time or add workers; the record of how each result was produced stays intact.
Rockie applies a hardware, runtime, and spend policy to each job, compares eligible capacity, launches the work, and records the run. Groups can route jobs to their own cloud accounts or lab hardware.
Set the hardware, time limit, and maximum spend before the experiment launches. The same policy travels with every retry and branch.
Hardware, runtime, and spend limits live in the same job specification, so there is one place to look and one thing to review.
Job status, logs, and provider-reported usage stay attached to the experiment that produced them, not scattered across dashboards.
Have GPU credits, a cluster allocation, or lab boxes? Wire them in and Rockie runs there first, with the same tracking and the same policies.
Rockie runs on the reasoning layer you already pay for: a Claude or ChatGPT subscription, or your own API keys for Anthropic, OpenAI, Google, Groq, DeepSeek, Kimi, Qwen, and more. Credentials stay encrypted and scoped to your own environment.
Drive Rockie's tools and skills from the command line, or hand off a whole job and let the cloud environment run it while your laptop sleeps.
needs a rockie account · rockielab.com
Rockie is built and run by Sam Larson and the Pebble team, and it powers every finding on the research page. If you run computational experiments for a living, bring your questions and your compute, and keep the record of how every answer was produced.
try rockie at rockielab.com → ← back to pebble