High-Assurance Systems,
SIMD Vector Cognition & Swarm Metaheuristics.
Explore the primary engineering divisions of HFLabs: in-memory GraphRAG & swarm optimization, heterogeneous ML & query acceleration, and bare-metal Intel Level Zero / Poplar native substrates.
Systems Topology & Architecture
Cognition & Swarm Intelligence
In-memory graph retrieval, SIMD vector search, and multi-algorithm swarm optimization. Keeps the entire graph resident in RAM to eliminate disk-bound lookups, powering agent reasoning through FastMCP and desktop visualization HUDs.
[Cognition & Swarm] Hypertensor GraphRAG: 2-hop traversal in 0.15 ms | SIMD 128d | FastMCP 65+ tools active
3 Specialized Divisions
High-assurance architectures spanning in-memory GraphRAG, autonomous swarm optimization, LevelZero.NET GPU acceleration, and bare-metal compute substrates.
Cognition & Swarm Intelligence
Vector & Swarm EnginesIn-memory graph retrieval, SIMD vector search, and multi-algorithm swarm optimization.
Core Projects & Runtimes
- Hypertensor GraphRAG Coming soon
In-memory typed property-graph store with hybrid dense and BM25 retrieval. Dense search uses SIMD TensorPrimitives, with optional Level Zero GPU offload. It gives agents multi-hop graph context through a 65+ tool FastMCP server and a desktop HUD.
HT GraphRAG Product Page Coming soon → - HyperSwarm Optimization Suite Production
Swarm optimization framework covering core, hybrid, advanced and multi-objective algorithms. It has Director-managed subswarm ensembles, convergence and diversity monitoring, Pareto-sorting threshold tuning, and a desktop visualizer.
HyperSwarm Product Page → - SwarmDev Production
Deterministic prompt-to-workflow pipeline. It does intent classification, planning, knowledge-graph construction, frame routing, policy-gated normalization and diagnostics. Optional kGNN reasoning and wavescan layers are opt-in.
SwarmDev Product Page → - SwarmBench Production
Benchmarking, characterization, elimination tournaments and real-time visualization HUD running 34 optimization algorithms against 110 test functions with HyperCache config reuse and self-validating LaTeX/JSON profiles.
SwarmBench Product Page → - Interactive PNN Decision Matrix Live Web App
Interactive diagram and browser-based visualization tool for probabilistic neural network decision matrices. Live on GitHub Pages.
Live Web App →
Machine Learning & Acceleration
ML & Inference EnginesAgentic AI building blocks, zero-allocation LLM runners, and IPU-offloaded query engines.
Core Projects & Runtimes
- ML.Agentic (v1.0.0) Production (v1.0.0)
21 zero-dependency multi-agent architectural building blocks, MCP (Model Context Protocol), A2A (Agent2Agent), and JSON-RPC 2.0 for .NET 10 with sub-millisecond execution and 0 external dependencies.
ML.Agentic Product Page → - l0llm (v2.0.0) Production (v2.0.0)
Ultra-fast, zero-allocation native LLM runner and OpenAI-compatible REST server in pure C# for Intel GPUs. Benchmarked at 41.3 tok/s on Iris Xe with zero GC allocations.
l0llm Showcase → - HFLabs Machine Learning Suite Production
Spiking neural networks, contextual and intergraph inference, and knowledge-graph embeddings (KGE). Hyperparameter optimization (Bayesian, Hyperband, BOHB, ASHA-BOHB, TuRBO) and graph-constrained beam decoding.
ML Suite Product Page → - Accelerator Production
Unified hardware-accelerator layer. A gRPC service executes plan envelopes across IPU (Poplar), CUDA, OpenCL, Level Zero, FPGA and DMA/RDMA backends with compiled-graph caching.
Accelerator Product Page → - IPU SQL Query Acceleration (MVP) MVP
A PostgreSQL extension and a SQL Server CLR bridge offload aggregates (COUNT, SUM, AVG, MIN, MAX) and vector distance ops (cosine, dot, L2) to the IPU through a gRPC microservice.
QueryAccel Product Page →
Hardware & Native Substrates
Bare Metal & InteropIntel Level Zero GPU interop, diagnostics, Graphcore Poplar bindings, and header-to-binding generation.
Core Projects & Runtimes
- LevelZero.NET (v2.0.0) Production (v2.0.0)
High-performance zero-allocation managed .NET 10 bindings for Intel oneAPI Level Zero runtime with pre-compiled SPIR-V and .zebin kernel catalogs across 11 Intel GPU architectures.
LevelZero.NET Product Page → - l0check (v2.0.0) Production (v2.0.0)
Fast standalone CLI diagnostic tool for inspecting Intel GPU architectures, driver capabilities, memory topologies, and validating compute kernel execution.
l0check CLI → - Poplar.NET (v1.0.0-preview) Production (v1.0.0-preview)
World first zero-allocation C# / .NET 10 runtime engine and resident graph orchestration SDK for Graphcore Colossus IPUs (Bow-2000, MK2, IPU-POD).
Poplar.NET Product Page → - Poplar Explorer Fork Production
Basic framework and interactive exploration workbench for rapid prototyping of Poplar programs on Graphcore IPUs.
Poplar.NET Product Page → - Abigen (in development) In Development
libclang-driven generator that reads C/C++ headers, emits reviewed/verified C# bindings, and manages upstream release tracking with fail-closed validation.
Abigen Product Page →