Local AI / Distributed Systems / 2026
Sylph
Local intelligence. Shared learning.
I am building a Rust runtime for local inference and exploring peer-to-peer exchange of model adapters.
Research — In development
Inspect an arriving adapter.
Inspect the compatibility and integrity gates around a received model-adapter update. Click a step, connection, or trace to inspect it.
Start the walkthrough, or inspect any step before running it.
Simplified from the project architecture. Runs locally; no live models, credentials, or external services.
The engineering
Inside
the decisions.
Connect local inference and shared learning through explicit trust boundaries.
Local AI needs more than a model loader. It needs a consistent runtime, a way to learn from useful work, and a trust boundary around knowledge received from other machines.
01The design decision
+
Separate the inference backend, adapter exchange, and validation layers. Check adapter compatibility and payload integrity before accepting a received update.
02The trade-off
+
Decentralized learning adds compatibility, privacy, and trust problems to inference. Module-level checks are useful evidence; broad model support and the complete learning loop still need end-to-end evaluation.
03Separating the model backend
+
Developed a backend-agnostic inference contract and runtime infrastructure in Rust.
04Exchanging compatible adapters
+
Explored task-driven learning and model-adapter workflows, including compatibility checks for exchanging adapter updates.
05Checking received payloads
+
Implemented peer-to-peer delta validation around base-model identity, adapter structure, payload integrity, and bounded decompression.
06Testing composition and deployment
+
Developed composition checks and adversarial tests for received adapters, with deployment tooling spanning local environments and Kubernetes/Helm.
↳Read the system flow
+
Sylph · Public explanation of the implementation boundaries.
- Local runtimeA backend contract separates inference from the surrounding system.
- Candidate adapterTask-driven learning explores deriving an adapter update from useful work.
- Exchange checksCheck base-model identity, structure, payload integrity, and decompression bounds.
- Evaluation boundaryCompatibility is a prerequisite; useful and safe model behavior needs separate evaluation.
↳Implemented and tested at module level
+
Runtime contracts, adapter compatibility, payload validation, composition checks, and adversarial tests exist in the private implementation. Local and Kubernetes/Helm tooling support deployment experiments.
↳End-to-end validation remains open
+
The complete learning loop and broad model support remain under evaluation. The browser exhibit runs SHA-256 and base-model checks on sample data; it does not execute the Rust runtime or evaluate model quality.
↳Trust boundary: peers and payloads
+
Received adapters cross an untrusted peer boundary. Identity, structure, digest verification, and bounded decompression address malformed or incompatible input. A matching digest establishes byte integrity, not a trustworthy sender.
↳Trust boundary: behavior and recovery
+
Poisoning, private-data leakage, unsafe composition, and rollback require separate evaluation. No general safety, privacy, or rollback guarantee is made for the current implementation.
Research — In development
Inspect
the work.
An independent project in development, with a public project overview and a private implementation. Current work explores inference, learning, adapter exchange, and validation. Broad model coverage and end-to-end safety remain goals requiring continued evaluation.
Next project: Memory-Arc ↗Source material
Public project overview