Making medical AI private by default
Enclara exists because patients shouldn't have to choose between AI-powered healthcare and data privacy. We believe these should be the same thing.
Privacy through math, not policy
Most privacy solutions are legal promises — terms of service, data processing agreements, compliance certifications. These are important, but they rely on trust. Enclara takes a different approach: the server cannot see patient data, not because of a policy, but because of the mathematics of Fully Homomorphic Encryption.
Cryptographic guarantee
Patient data is encrypted with a secret key that never leaves their device. Without it, the data is computationally indistinguishable from random noise.
Verifiable privacy
Our entire pipeline is open source. Anyone can inspect the code, verify that no plaintext data is transmitted, and audit the FHE circuit.
Practical utility
We chose skin cancer classification because it’s a real clinical need with severe privacy implications — this isn’t a toy demo.
Built by founders who ship, not just theorise
Most teams in this space publish papers. We build products. Edward's selection for 0xPARC — the premier applied cryptography program at the frontier of FHE, ZK, and MPC — gives us mastery of the cryptographic primitives. But what sets us apart is the engineering firepower to take that knowledge from concept to production. Hardware-level design, full-stack systems, and relentless shipping speed — this is the team turning privacy-preserving AI from a dream into deployed reality.
Edward Wei
Co-FounderMEng General Engineering at the University of Cambridge (Churchill College). Churchill Scholar, Bill Brown Prize and Kevin Knowles Prize winner. ASIC design intern at Renesas Electronics, working with Cadence tools on analog circuit design and SystemVerilog/UVM verification. Research intern at the Cambridge Division of Electrical Engineering, co-authoring work on nanoparticle optical filters. 0xPARC Summer 2025 alumnus — selected for the premier applied cryptography research program at the frontier of FHE, zero-knowledge proofs, and multi-party computation. Trained for the Asian Physics Olympiad through the Hong Kong Academy for Gifted Education.
LinkedIn →Leo Hammett
Co-FounderMEng Manufacturing Engineering at the University of Cambridge (Sidney Sussex College). Full-stack engineer across Python, TypeScript, C++, and Swift with production experience at AbbVie Pharmaceuticals as an AI Data Engineer. Multiple first-place hackathon finishes including Tezos and Polkadot, with over $6,500 in prizes. Arkwright Engineering Scholar.
CV →What we've built
FHE-compatible neural network
A 5-bit quantized VGG11 variant (QuantVGG11Patch) trained on HAM10000 and compiled to an FHE circuit via Concrete-ML and Brevitas. Classifies skin lesions into 7 diagnostic categories entirely under encryption.
Patch-based inference system
224×224 images split into a 7×7 grid of 32×32 patches. Each patch is independently encrypted, transmitted, and classified. Client-side aggregation combines the 49 results into a final diagnosis.
iOS client application
A SwiftUI app that handles key generation, image capture and cropping, patch encryption via PythonKit bridge to Concrete-ML, and result decryption. Keys are generated once and reused across scans.
Encrypted inference server
An AWS EC2 backend that stores evaluation keys and runs the FHE circuit on encrypted patches. It processes data it cannot read and returns results it cannot interpret.
Where this goes
Skin cancer classification is the first application. The underlying architecture — encrypted neural network inference on medical images — generalizes to any diagnostic imaging domain where patient privacy is critical.
Beyond dermatology
The same FHE inference pipeline can be adapted for radiology, ophthalmology, pathology, and other imaging-based specialties where patients need AI assistance without data exposure.
Regulatory advantage
When the server mathematically cannot access patient data, GDPR and HIPAA compliance becomes architecturally simpler. Data protection is built into the system, not bolted on.
Enterprise deployment
Hospitals and clinics can use AI diagnostic tools without building their own ML infrastructure or risking patient data in third-party clouds.
Performance optimization
As FHE libraries and hardware accelerators improve, inference times will drop. The architecture is ready — we’re limited by FHE compute speed, not design.
Open source. Open to collaboration.
Enclara is fully open source. If you're working on FHE, medical AI, or privacy-preserving computation, we'd love to hear from you.
View on GitHub