There’s a faster GPU hiding inside yours. You’ve already paid for it.
Eppy is a hunch engine with the instincts of an expert performance team. It finds it. Proves it. You keep it.
More users per GPU, quality checked.
Public benchmarks, same GPU, same promise to users, answers checked.
Found automatically.
Answers checked.
2.5×Best known setup in under an hour.
Quality unchanged. Busy chat: 4.6×.
7.06Found automatically,
on documents it had never seen.
5.35measured on held-out traffic Eppy never saw · same response promises · H100 cards: requests per second on one GPU
Estimate based on our public benchmark; your result depends on your model and traffic.
No vLLM expert needed.
Found automatically, on new documents.
Chat needs a different recipe.
vLLM · default settings
Your GPUs are next.
Send us your model and a slice of real traffic. We send back your number.
- No changes to your model
- Same response times
- Answers checked
Anything that's expensive to test.
Eppy isn't specific to GPUs. It works on any system where each test costs real time or money. Logistics is live. More are in the oven.
Our methods are out in the wild.
Running inside some of the world's most established labs, from CERN to ETH Zürich.
all 19 sources
- Ray · 2021 · Built into Ray Tune, a widely used tuning library
- Continental · 2022 · Tuned how cars' cameras spot their own detection failures
- U Toronto · 2022 · A planner in the Olympus self-driving-lab toolkit
- UPC Barcelona · 2022 · Calibrated a real building's heating and cooling model (45 settings)
- CERN · ATLAS · 2023 · Tuned the AI that hunts for a new particle at the Large Hadron Collider
- ETH Zürich · 2023 · Designed a walking robot's new knee: 33% less motor effort, tested on the real ANYmal robot
- Google · 2023 · Rebuilt its model inside Vizier, Google's optimisation library
- Volkswagen · 2023 · Tuned AI models on vehicle data, across three studies
- DLR · 2023 · Built into the German Aerospace Center's system-architecture design toolkit
- NVIDIA · 2024 · Built it into their chip logic-synthesis system
- Intel Labs · 2024 · Uses it to stress-test large language models (LLMart)
- ZTE · 2024 · Designed an ultra-wideband antenna
- Optuna · 2024 · Available as a sampler on OptunaHub
- Ant Group · 2024 · Tuned LLM serving engines: up to 68% more requests than default settings, now in production
- MIT-IBM Watson AI Lab · 2025 · Used it to teach an AI model how to optimise
- MIT · CRESt · 2025 · Built into a robotic electrochemistry lab
- Sony AI · 2026 · Tuned a robot arm's table-tennis serves, played against professionals
- fal · 2026 · Tuned the training of a 22B-parameter video model
- SJTU · Zhejiang Lab · 2026 · Designed DNA binders: 2–3× as many strong ones as the standard lab method, confirmed in the lab
Of a problem: hard to evaluate.Checking whether an answer is good takes a lab, a fab, a solver or a week.
Named in homage to NP-hard, computer science's classic name for hard problems. Same spirit, different bottleneck: here the hard part is the test.
Come build Eppy.
We're building an AI that gets more out of expensive systems than the experts do. It's early, it's hard, and your work shows up in real numbers.
What you'd touch in week one
The agent behind 2.5× more users per GPUsame GPU, answers checked
The agent that made delivery routes shorterabout 8% shorter than an expert setup
Methods already running at CERNand in teams across industry