Sarthak Pati
Founder of VerySafe.ai
AI Safety Researcher & Engineer · Vice Chair, MLCommons Medical Working Group
I build AI for organizations where a wrong answer costs more than a headline: hospitals, regulated industries, anywhere the output has to be trusted before anyone can act on it. I hold a Ph.D. in Computer Science from the Technical University of Munich (summa cum laude) and have spent the last 15+ years taking AI from prototype to production.
Most of that work now happens at VerySafe.ai, where I'm building SafeCompute, a policy-aware compute platform that attaches cryptographic proof to every AI model run, using remote attestation, supply-chain provenance, and signed audit lineage. The point is that you can run frontier and open-source models in the places where someone will eventually ask you to prove what happened.
I also run Vaiyu Solutions, where the work is other people's AI: architecture through to production, usually on a pilot that has stalled somewhere short of launch.
Along the way I've led $9M+ in NIH/NCI-funded research, published in Nature, Nature Communications, and IEEE Transactions on Medical Imaging, and I serve as Vice Chair for Algorithmic Development at the MLCommons Medical Working Group.
I think open software makes for better science, so most of my code lives in the open.
What I build
From research to production
Four things people hire me for. Each one has had to work outside a paper, in production, under someone else's rules.
End-to-end AI systems
I take AI from first prototype to clinical-grade production, multimodal data and all.
GaNDLF: 30% faster prototyping, now an MLCommons project
Confidential & federated compute
Models train and benchmark across institutions while the sensitive data stays put.
$9M+ in NIH/NCI grants led · deployed on 6 continents
Optimization & deployment
I get models running where compute is scarce and latency budgets are real, from edge hardware to HPC.
10–50% less compute · up to 70% lower inference latency
Benchmarking & evaluation
Evaluation of medical and enterprise AI that other people can actually reproduce and check.
MedPerf: federated benchmarking across institutions
Open source
Selected projects
Editor's Choice, Communications Engineering (Nature)
GaNDLF
Do-It-Yourself Deep Learning framework for everyone: low-code AI pipelines for healthcare.
MLCommons Working Group
MedPerf
Open platform for federated benchmarking of medical AI models across institutions.
Nature Communications
FeTS
Federated Tumor Segmentation: largest real-world federated learning study (71 sites, 6 continents).
securefederatedai
OpenFL
Open-source federated learning framework for healthcare and life sciences.
CBICA / UPenn
CaPTk
Cross-platform toolkit for medical image processing and analysis.
arXiv:2410.00173
GaNDLF-Synth
Synthetic data generation for medical imaging: training AI with artificial data.
Trying to ship AI in a regulated domain?
That's my favorite kind of problem. The first call is on me.
sarthak@verysafe.ai