Research on private inference
Today, a prompt sent to an AI model can usually be read by the service you use, any router in between and the cloud provider running the model.
We support research into reducing that exposure: running AI on sensitive data while limiting what providers, intermediaries and infrastructure can see. Each approach is compared on what it protects, who must still be trusted, and what it costs in speed and model quality.
Research lines
Multi-party computation
Splitting computation across parties so that no single one can read the data, including on standard GPUs.
Hardware isolation
Trusted execution environments and remote attestation, with the trust placed in vendors and firmware written down.
Anonymised API access
Redaction, pseudonymisation and relay designs, tested for what identity and metadata still leak.
Encrypted inference
Fully homomorphic encryption prototypes that measure latency, model constraints and output quality.
On-premises and local inference
Keeping the most sensitive data on your own hardware, and partitioning what goes to the cloud.
Local Inference News. A daily briefing on running AI on your own hardware: new open models, inference engines, local rigs and what a task costs compared with closed models.
Open Local Inference NewsResearch project: IPRC. An early-stage project exploring private inference on encoded data.
Visit iprc.aiDirectories. 76 European machine-learning research groups and funders, and 20 cryptography audit firms with 170 of their key public works, each linked to its evidence.
Browse the directoriesFour practices, one standard
Each practice answers a different question. All of them end the same way: with work the responsible person can inspect.
-
Regulatory and market intelligence monitoring
We run monitoring processes that track local, regional and national government sources across Spain and Europe, alongside market signals, and turn them into short reports your analysts can use straight away.
Each report cites its source, date, status and confidence, and is reviewed by one of our analysts before it goes out. It supports your analysts’ judgment; it doesn’t replace it.
-
Private inference research
We study how to run AI without exposing sensitive inputs, identities or results, and we test each approach against a written threat model.
We compare trusted execution environments, anonymised API access, multi-party computation, fully homomorphic encryption and on-premises inference with data partitioning: what each protects, who must still be trusted, and what it costs in speed and model quality.
-
Private agent workflows
We turn a recurring task into an AI-assisted workflow your team owns: documented, permissioned and evaluated before it goes live.
Work usually includes a workflow and data-flow map, an evaluation set, a pilot and an operating handover. Review stays with the people who know the work.
-
Investment in private inference
We invest up to €250,000 in EU researchers with novel ideas for private inference tools, especially tools that mix local and private cloud inference.
We start with the work: the problem, the customer, the technical approach, evidence so far, the team and the next milestone. Investment decisions are kept separate from paid client work.
Every engagement ends in something you can check
Whether it’s a workflow, a research question or an investment, the work follows the same four steps.
-
Frame
Write down the question, the scope, the assumptions and who owns the decision.
-
Build
Develop the workflow, experiment, model or diligence view.
-
Check
Test sources, permissions, failure modes and limits, and record what couldn’t be verified.
-
Hand over
Put the evidence in front of the person who decides, in a form they can inspect.
Investment thesis
We back researchers in the EU who are turning private inference into tools people can use. Tickets of up to €250,000.
The problem
Most AI runs on someone else’s computer. Every prompt, file and result can be read by the service, the routers in between and the cloud provider running the model. For hospitals, law firms, banks and public bodies, that exposure is why their most sensitive work still stays away from AI.
No single technique solves it at an acceptable cost. Local models are private but limited in size. The cloud is powerful but has to be trusted. The practical answer is a mix: keep the most sensitive data on local or on-premises hardware, and send only what is needed to a private cloud that cannot read it.
What we invest in
Local and private cloud, together
Inference split between devices or on-premises servers and a private cloud, with data partitioned by how sensitive it is.
Cryptography and hardware, combined
Multi-party computation and trusted execution environments used together, so no single operator can read the data while speed stays close to ordinary cloud inference.
Tools, not only protocols
SDKs, runtimes, routers, attestation and evaluation tools that let engineering teams adopt private inference without a cryptography team.
Claims that can be checked
Threat models, reproducible benchmarks and open verification, so a customer can test a privacy claim instead of taking it on trust.
- Ticket size
- Up to €250,000 per team.
- Who we back
- Researchers and research-led founders based in the EU.
- Focus
- Private inference, and tools that mix local and private cloud inference.
- Stage
- From research prototype to first product. Investment decisions are kept separate from paid client work.
Send us your deck
A short deck is enough: the problem, your approach and threat model, evidence so far, the team, and what €250,000 would let you prove.
Straight answers
Short answers to what people usually ask before a first conversation.
What does “private” mean in your work?
Before we call anything private, we agree in writing how data, inference, storage, logs, access, retention and outside providers are handled. Where a guarantee depends on trusting someone, we say who.
What is IPRC?
IPRC is an early-stage research project on private inference that we support. It isn’t a product yet. You can follow it at iprc.ai.
Does a regulatory brief replace legal advice?
No. A brief gives cited signals, status, provenance and context. The client, law firm or in-house team keeps the final legal and business judgment.
How do research and client work relate?
Technical know-how moves between practices. Client and founder information doesn’t: it stays within the agreement it was shared under and is never reused without permission.
What should I bring to a first conversation?
The question or workflow, the people responsible, the sources or systems involved, the decision you need to make, and any constraints on privacy, review or publication.
Join the lab
We hire engineers and researchers only. Every role works on the problem our research is built around: partitioning data between local hardware and private or public cloud inference, so that nobody in between can read it.
Private inference research engineering
Design and benchmark inference that combines multi-party computation with trusted execution environments on standard GPUs. Write the threat models and the reproducible benchmarks behind every claim.
Data partitioning and local inference
Build the layer that decides which data stays on devices or on-premises servers and what goes to private or public cloud models, and measure what leaks and what it costs in quality and speed.
SDK and open-source tools
Build open SDKs, runtimes, routers and attestation tools so engineering teams can adopt private inference without a cryptography team of their own.
Engineering for analysts and consultants
Build and run the tools our analyst and consulting team relies on: source monitoring pipelines, evaluation sets, permission-aware integrations and handover to clients.
No role that fits exactly? If you have built something in private inference, cryptography or systems, send a repository, paper or benchmark you can talk through.
Submit your open applicationBring us a question worth checking.
Clients
Request a demoDemos of regulatory and market intelligence monitoring, and workflow projects.