Home 01 Technology 02 Projects 03 Research 04 About 05 GitHub ↗ Contact FuryLogic

FuryLogic Open Research

Research, published in the open.

FuryLogic Open Research is the publication arm of the company. It holds the papers we have deposited with a DOI, the research questions we are actively working on, and the limits of what we can currently claim.

01 — How this record works

A paper is a paper. Not a repository.

Research software exists to support a claim. We publish the claim, the method and the result as a citable document — and we do not dress a prototype up as a product to make the record look larger.

Publication

DOI first

Every paper on this page has a resolvable DOI. The citation remains valid even if a repository is moved, renamed or archived.

Evidence

Named baselines

Results state the machine, the model file, the flags and the comparison point. “Faster” is meaningless without the thing it is faster than.

Honesty

Limits included

Where a result does not generalise, or a mechanism does not work yet, the paper says so. A gap audit is part of the publication, not an embarrassment.

02 — Publications

Published preprints.

Four papers, deposited on Zenodo, maintained by FuryLogic Labs. Each entry links the DOI and, where available, the paper PDF.

Paper 01 2026
DOI 10.5281/zenodo.21471506

OpusEdge: Telemetry-Guided Dynamic Compute Allocation for Dense, MoE, and Hybrid SSM-Attention Architectures

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.

Irfan Mahir Zenodo preprint LLM inferenceKV cacheSparse attention
Paper 02 2026
DOI 10.5281/zenodo.21549808

VUGVA: A Software-Defined Virtual Unified GPU VRAM Architecture with CPU-Bypass Hybrid Memory for Non-NVLink Multi-GPU Clusters

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.

Irfan Mahir Zenodo preprint GPU virtualizationMemory tieringCPU bypass
Paper 03 2026
DOI 10.5281/zenodo.22810393

Attention Stream Cache: Reducing DRAM Bandwidth Saturation in LLM Attention Layers

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.

Irfan Mahir Zenodo preprint DRAM bandwidthPrefetchingCUDA
Paper 04 2026
DOI 10.5281/zenodo.21756604

Synthetic Intelligence: A Graph-Physics Substrate for Structural Reasoning

A research programme investigating reasoning without gradients, backpropagation, learned embeddings or next-token prediction. Concepts are cavities with mass, displacement and velocity; inference is the propagation of local forces through coupled relations rather than a sampling pass over token probabilities.

Irfan Mahir Zenodo preprint Synthetic intelligenceGraph physicsReasoning

Early-stage research programme. Openly documented, including its gaps.

Citation format: Mahir, I. (2026). Title. Zenodo. https://doi.org/10.5281/zenodo.…

03 — Research interests

Questions we are actively working on.

These are research directions, not product lines. Some have published results; others are open problems where the honest answer is still unknown.

01

LLM inference optimization

Where time and memory actually go during prefill and decode: batched execution, KV cache residency, prefix reuse, speculative decoding and the scheduling between them.

OpusEdgeAttention Stream Cache
02

Model architectures

Dense transformers, mixture-of-experts routing and hybrid state-space attention, studied as families with different compute and memory behaviour rather than one uniform target.

OpusEdge
03

Distributed inference

Serving a model across devices that do not share a fast interconnect: memory pooling, expert placement, peer transport and graceful degradation when a node is slow.

VUGVA
04

GPU and CPU acceleration

Kernel-level work on both sides of the bus: quantized matrix multiplication, DMA descriptors, prefetch scheduling, SIMD, and the cost of moving data rather than computing on it.

VUGVAAttention Stream Cache
05

Synthetic intelligence

A separate research lineage: structural reasoning over a graph-physics substrate, decision readout from settled fields, and systems that defer instead of guessing.

Synthetic Intelligence
06

Search and retrieval infrastructure

Hybrid lexical and dense retrieval, rank fusion, cross-encoder reranking, and evaluation that can show a retrieval change helping or hurting instead of asserting it.

07

Systems engineering

Databases, kernels, runtimes and protocol surfaces — the layer below the product, where consistency models, memory safety and boot time are the features.

04 — Method

How a result becomes a paper.

The process is deliberately unglamorous. It exists so a reader can check the work without asking us for permission.

  1. Measure first

    Every claim starts as a number from a named machine and a named model. Where a result does not hold, it is published as a limitation rather than dropped.

  2. State the mechanism

    A result is only useful if the reason is legible. Papers describe the mechanism precisely enough to be reimplemented or falsified.

  3. Reproduce in code

    Reference implementations accompany the papers so a reader can run the primitive rather than trust the plot.

  4. Keep the record

    Preprints are deposited with a DOI and remain citable. Corrections supersede earlier versions instead of quietly replacing them.

On status: Synthetic Intelligence is an early-stage research programme. Its paper is public, its gap audit is public, and it is not offered as a working general-purpose system. We would rather state that plainly than let a reader discover it later.

Engage with the work

Cite it, reproduce it, or challenge it.

We are interested in correspondence that engages with the mechanism: a reproduction that disagrees, a baseline we missed, or a constraint that changes the conclusion.