Applied research

Frontier research on foundational model training, private AI deployment, and enterprise-grade security for sovereign organizations.

Learn more

Privacy

Confidential computing

We build frameworks that convert standard transformer architectures into their private counterparts. These frameworks enable inference and fine-tuning while preserving privacy for data and model owners.

Performance

Novel permutation and factorization techniques for privacy-preserving AI with minimal computational overhead.

Performance

Novel permutation and factorization techniques for privacy-preserving AI with minimal computational overhead.

Performance

Novel permutation and factorization techniques for privacy-preserving AI with minimal computational overhead.

Security guarantee

Secure multi-party computation to guarantee privacy across distributed datasets.

Security guarantee

Secure multi-party computation to guarantee privacy across distributed datasets.

Security guarantee

Secure multi-party computation to guarantee privacy across distributed datasets.

Data protection

Differential privacy integration to mathematically prevent data leakage and model memorization.

Data protection

Differential privacy integration to mathematically prevent data leakage and model memorization.

Data protection

Differential privacy integration to mathematically prevent data leakage and model memorization.

Trade-off

Fine-grained privacy controls to perfectly balance cryptographic security with inference speed.

Trade-off

Fine-grained privacy controls to perfectly balance cryptographic security with inference speed.

Trade-off

Fine-grained privacy controls to perfectly balance cryptographic security with inference speed.

Specialization

Foundational model training

We build the architectural primitives that power next-generation AI systems. From pre-training dynamics and data curation to scaling laws and emergent capabilities.

Data

Scalable data curation pipelines for high-signal, diverse pre-training corpora.

Data

Scalable data curation pipelines for high-signal, diverse pre-training corpora.

Data

Scalable data curation pipelines for high-signal, diverse pre-training corpora.

Compute

Adaptive compute allocation to balance model depth, width, and training efficiency.

Compute

Adaptive compute allocation to balance model depth, width, and training efficiency.

Compute

Adaptive compute allocation to balance model depth, width, and training efficiency.

Evaluation

Emergent capability benchmarking for systematic evaluation across capability thresholds.

Evaluation

Emergent capability benchmarking for systematic evaluation across capability thresholds.

Evaluation

Emergent capability benchmarking for systematic evaluation across capability thresholds.

Frameworks

Architecture-agnostic training frameworks with modular, configurable training objectives.

Frameworks

Architecture-agnostic training frameworks with modular, configurable training objectives.

Frameworks

Architecture-agnostic training frameworks with modular, configurable training objectives.

Research in production

Confidential speech-to-text

Sotto puts our confidential compute research to daily use: a free macOS menu bar app for push-to-talk dictation. Hold Right ⌘ to speak, release to transcribe.

Enclave API

Audio is transcribed by a Deepgram model running inside Prem's hardware-encrypted Enclave. No one outside can read it, including us.

Encrypted end to end

Encrypted end to end. Audio travels encrypted from your Mac to the enclave and back. Nothing is stored. Sub-200ms latency. Confidential compute at dictation speed.

Try Sotto

Research by

The state of private LLM inference

We surveyed the landscape of privacy-preserving inference for Transformer-based LLMs. Discover our comparative analysis of cryptographic techniques, performance trade-offs, and practical deployability for real-world enterprise architectures.

Read the paper