Artificial Intelligence & Machine Learning
Model systems built to be measured, not merely demonstrated.
Independent engineering company · AI · HPC · Systems
FuryLogic is an independent engineering company exploring the boundaries of artificial intelligence, high-performance computing, and intelligent software infrastructure.
Measured in each project’s own benchmark suite. Sources are named on the project pages.
01 — The company
FuryLogic designs, writes and operates the infrastructure that intelligent software stands on — from CUDA kernels and storage engines to retrieval systems and the protocols between them.
FuryLogic is an independent technology and AI engineering company. Our work sits where performance is decided: inside inference engines, memory hierarchies, search pipelines and system runtimes. We are not a platform vendor, and we do not sell access to a model. We build software, publish the research behind it, and run it.
The portfolio is deliberately narrow in scope and deep in implementation. One team owns each system from the kernel to the HTTP surface, which means the person who wrote the scheduler is also the person who has to explain its tail latency. That constraint is the method.
We measure against named baselines on hardware people actually own — a single 8 GB GPU, a CPU-only server, a router with 4 MB of flash — because a result that requires a datacenter is not a result most operators can use.
“Engineering intelligence is not a slogan. It is the discipline of making a system explain itself.”
FuryLogic — engineering principles02 — Core disciplines
The categories overlap on purpose. An inference engine is a systems problem; a search engine is a distributed-systems problem; both are judged by measurement.
Model systems built to be measured, not merely demonstrated.
The work between a trained model and a served token.
Memory hierarchies treated as the design constraint.
Coordination that degrades predictably.
Retrieval you can evaluate and therefore improve.
Kernels, runtimes and storage engines in Rust and C.
03 — Featured systems
Four systems in active development. Each one solves a problem we hit ourselves, and each one is measured against the tool it replaces.
A pure-Rust inference engine with its own CUDA kernels. Batched decode across model families, expert placement for MoE models larger than VRAM, prefix caching for agent traffic, and speculative decoding.
One engine. Every data model. MVCC and WAL storage, HNSW vector search with INT8 plus f32 rerank, a cost-based RAG planner, bi-temporal agent memory, and RESP wire compatibility.
Self-hosted search for AI agents. Hybrid ranking, query-aware snippets, near-duplicate collapse, per-language full-text indexes, and measurable retrieval evaluation with nDCG, MRR and recall.
An adaptive fetch gateway for the hostile web. TLS fingerprint impersonation, per-domain strategy learning, session-cookie reuse, browser rendering only when cheaper tiers fail, and clean DOM-to-markdown extraction.
04 — Engineering philosophy
We do not separate the engineering from the claim. A system that cannot be measured cannot be defended, and a benchmark that cannot be reproduced is a story.
A number without a machine, a model and a baseline is a marketing claim. Benchmarks in this portfolio name the GPU, the CPU, the model file and the flags used to produce them.
When an existing component cannot meet the requirement, we build the component. That is why this portfolio contains a database, an inference engine, a search stack and a memory manager rather than wrappers around them.
Systems fail in ways their operators must understand. We prefer explicit refusals, typed outputs and honest gap audits over confident silence or a fallback that hides the problem.
A kernel without a scheduler, an engine without a protocol and a paper without a reproducible path are incomplete. We carry work from the lowest layer to the interface a person actually uses.
Build the thing. Then prove it works.
Read about FuryLogic05 — FuryLogic Open Research
Papers are deposited with a DOI and stay citable. We link the publication, not a repository — the code that exists to support a paper belongs with the paper.
A single per-token importance signal, Δ, is extracted from dense hidden-state drift, native SSM selectivity or router-softmax entropy, then used to drive KV cache eviction, sparse attention, head gating and state compression without retraining. Measured at 93.8 % KV cache reduction at 65,536 tokens on one 8 GB GPU across dense, MoE and hybrid model families.
GPU memory is fragmented across cards, and frameworks that spill through the CPU lose a large share of throughput. VUGVA virtualizes multi-GPU VRAM into one pool, transfers with 64-byte DMA descriptors so the CPU never touches payloads, and tiers memory from VRAM to DRAM to SSD with automatic promotion and demotion.
Attention’s scattered KV cache reads carry an 80–84 % bandwidth penalty relative to sequential reads. ASC pairs a predictive attention router with an asynchronous streaming prefetch queue, staging the blocks a kernel will read next into L2 before it asks for them.
Four preprints on Zenodo · maintained by FuryLogic Labs
Enter Open Research06 — Open source
Most of this portfolio is public: Rust, C, C++ and Python under MIT, Apache-2.0 and AGPL-3.0. Canonical repositories live on our own infrastructure, mirrored to GitHub.
Start a conversation
We work on inference, infrastructure, retrieval and systems problems. If the constraint is performance, correctness or memory, that is the conversation.