Project · msearch
msearch
an intelligence runtime — composable pipelines for search, inference, training, and agents
A DAG execution engine for composable intelligence pipelines. Search, infer, traverse, train, coordinate — composed into a single execution plan, parallel across CPU and GPU, zero-copy throughout. Profiled for Apple Silicon's unified memory, deployable on any server.
Intelligence pipelines as a first-class intelligence runtime.
The dominant pattern for AI applications today is to chain separate services — search here, inference there, coordination somewhere else — across the network. As pipelines become more adaptive and agents compose their own plans at runtime, that architecture becomes the bottleneck.
msearch flips that. Pipeline are composed as a directed graph that the engine executes. Pipelines are made of up common intelligence primitives — vector search, inference, training, reasoning, indexing. Execution occurs across GPU and CPU cores in parallel, each operation comsumping the compute framework best suited to it.
The result is a class of application that hasn't really existed: a dynamic, on-device intelligence runtime — with latency budgets measured in milliseconds and privacy guarantees that hold because the data never left the device.
The architecture in one read.
Composable DAG execution
Pipelines are graphs of operators, not chained service calls. Search, infer, traverse, train, coordinate — composed and optimised together as one execution plan.
Zero-copy across the pipeline
Data stays in place from query to result. No intermediate copies, no serialisation between stages. The pipeline never duplicates what's already in memory.
Parallel CPU and GPU lanes
Operators choose where they run. Graph traversal stays on CPU; distance kernels and inference lift onto GPU. Both lanes coordinate through shared memory without copies.
Profiled for Apple Silicon
On unified memory, the zero-copy architecture extends all the way to GPU compute. No transfer tax between processors. A single binary that also deploys on any server.
Built on mgraph
mgraph carries structured state across environments. msearch turns that structured state into pipelines that produce structured output. The two are co-designed.
Where it goes from here.
- Pipeline runtime over UMAdevelopedOperator graph executor on Apple Silicon, zero-copy between stages.
- First-class vector indexdevelopedGPU-accelerated index built into the runtime — the vector DB project.
- Reranker and structured output operatorsdevelopedThe pipeline can return JSON-shaped, schema-validated results.
- Public open-core releasein flightSource, spec, examples, and the first wave of partner integrations.
- Graph-traversal operatorin flightTyped graph queries as a native operator alongside vector and rerank.
- Cross-platform runtimenextLinux + CUDA backend that mirrors the UMA contract for non-Apple GPUs.
- Distributed pipeline coordinationnextPipelines that span device, edge, and cloud as one execution plan.
- Specialised model operatorsnextNative operators for graph-neural-network reasoning over typed knowledge.
Tell us what you're working on.
We reply within two business days. If a call would be faster, book a thirty minute conversation.