01
Reliable AI systems
Measuring where serving stacks, deployment changes, and shifting context alter model behavior—and surfacing those risks before they become production failures.
Independent research groupEst. 2026
Research,
without the institution.
Enneper AI is a home for independent researchers pursuing rigorous work in AI systems, mathematical structure, and scientific discovery.
Fig. 01Enneper surface
01 / Purpose
Why Enneper
A serious idea should not need an institutional address to exist.
Independent researchers often have the skill and the question, but not the surrounding structure: collaborators, compute strategy, experimental critique, or a path from prototype to public artifact.
Enneper AI exists to help close that gap. We form small, focused collaborations around work that can be tested, reproduced, and shared. The goal is not volume. It is research with a clear claim and evidence strong enough to carry it.
02 / Research directions
What we work on
Following intelligence through the systems that make it possible.
Our agenda is deliberately narrow enough to test and broad enough to connect hardware, software, model behavior, and scientific use.
02
Hardware & numerical behavior
Understanding how accelerator arithmetic, precision, kernels, and portability shape correctness, performance, and reproducibility.
03
AI for scientific discovery
Applying long-context and systems methods to domains where better representations and reliable computation can expose new scientific structure.
04
Independent research infrastructure
Developing practical patterns for solo researchers: scoped collaborations, reproducible artifacts, compute-aware experiments, and public technical communication.
03 / Starting point
Research in practice
Built from work across ML systems, accelerators, and evaluation.
Enneper AI is initiated by Gianluigi Vitale, an ML systems researcher and systems engineer. His work includes DriftBench, presented orally at MLSys 2026, TPU kernel and numerical-characterization research, and artifact evaluation for MLSys and SOSP.
View the research profile- 236,985
- prompt-response pairs in DriftBench
- 105
- measured system configurations
- v4–v6e
- TPU generations in active research
01Ask a precise question.
02Build the smallest credible test.
03Make the evidence inspectable.
04Share what survives.
04 / Open call
For independent researchers
Bring the question you cannot stop thinking about.
You may be between institutions, working after hours, changing fields, or simply choosing to work independently. We care less about affiliation than whether you can state the problem clearly and show how you would begin.
This is likely a fit if you:
- have a concrete research question, not only a general interest;
- can contribute technically, experimentally, or through domain expertise;
- value reproducible work and honest negative results;
- want a focused collaboration, not a credential or a title.