SKiMet
A distributed metagenomic classifier built on top of the SMARTEn framework and integrating SKiM’s compressed index into a server/client architecture.
Why it exists
The motivation behind SKiMet is privacy-preserving classification. Sequencing data is sensitive: hospitals, field clinics, and personal genomics workflows can’t always send raw reads to a centralized service. SKiMet is the foundation for a setup where the classification service holds the reference index but never sees raw reads in plaintext, while the client never has to host a 17 GB reference DB locally.
My contribution
I built the gRPC-based server/client framework: the wire-protocol contract, the connection lifecycle, the streaming interface for sending reads to the server and receiving classifications back, and the server-side pipeline that routes incoming requests through SKiM.
I have not worked on the privacy-preserving cryptographic layer itself (homomorphic encryption, secure multiparty computation, or similar). That’s the next phase of the project; SKiMet today is the network framework that the privacy layer will eventually live on top of.
What it does today
A client connects to the SKiMet server over gRPC and streams sequencing reads. The server holds the SKiM compressed index in memory and runs classification through an Intel TBB pipeline that overlaps I/O, decompression, and lookup so a steady stream of client requests doesn’t stall on any one stage. Classifications stream back to the client as they’re produced.
The point is to decouple the classifier from the device that has the sequencing data: the server holds the reference index, the client holds the reads, and they exchange just enough over gRPC for classification to happen. That makes SKiMet useful anywhere a client machine can’t reasonably host a 17 GB reference DB locally.
Stack
C++23, gRPC, Intel TBB, SKiM (as the index backend), CMake. Cross-platform.
Status
In active development. The server/client framework and TBB pipeline work end-to-end against the SKiM backend. The privacy-preserving layer is open work, not yet started.