A private research lab working at the edge of data compression, neuromorphic vision and computational geometry. One recurring principle, studied across many domains: the shape of the part already knows the shape of the whole.
Every project here starts from a single observation — that in many systems the local already contains the global. Compression, vision, signal, form. When you find the right axis of symmetry, the data stops being data. It becomes pattern. And patterns carry themselves.
A lossless codec for neuromorphic event-camera streams. Treats sparse asynchronous events as a self-referential structure, and compresses them accordingly. Python wheels and a WebAssembly browser demo are published; a preprint is archived on Zenodo. US provisional filed.
A spatial container format for drawings, diagrams and layered geometric data. Progressive delivery, browser-native viewer, small footprint. Built for architecture, PCB, medical imaging — anything hierarchical that needs to travel through a narrow pipe.
A compression library for neural-network weights. Treats each parameter group as a branching decision tree to squeeze large language models into places they currently don't fit — smaller devices, narrower memory, tighter latency budgets — without the usual accuracy cliff.
A transport layer tuned for progressive, quality-aware content delivery. Content arrives in meaningful layers — coarse first, refined later — so that every byte already says something useful about the whole. Target workloads: live medical imaging, engineering drawings, map tiles.
An approximate retrieval engine for noisy, partial or imprecise queries. Same recursive idea as the rest of the lab — build the lookup path out of self-similar branches so that imperfect input still lands close to the right neighbourhood.
We publish. Preprints land on open archives so that the work is dated, readable, and independently verifiable. Binary distributions ship through the usual package registries. A provisional patent is on file for the method underlying the codec; commercial licensing is available directly.
See all researchCollaborations, licensing, residencies, technical questions — all of it lands in the same inbox. We usually reply within a few days.
Write to the lab"Every structure carries a smaller version of itself. We build the tools that listen for it."— Studio principle, MMXXVI
A private research lab exploring one recurring principle — self-similarity — across compression, vision, and computational geometry. One thesis. Many experiments. A slow, deliberate build.
We don't chase ratios. We look for the axis along which data folds onto itself — and let the folding do the work. Ratios are a consequence, not a target.
Experiments first, tools second, products third. Nothing ships until a bit-exact round-trip passes. Nothing becomes a product until it has at least one real user who pays for it.
Preprints public, binaries shipped, patents filed. The internals that matter stay internal. The parts that help people evaluate the work go out into the open.
Compression. Neuromorphic vision. Sparse geometry. Neural-network weights. Approximate search. Surface level they look unrelated; underneath there is only one pattern.
A theorem is worth nothing until it is code someone can run. A benchmark is worth nothing until it is reproducible on a laptop. Theory ships as software or it doesn't ship.
Every claim we publish comes with a way to check it. No trust-us numbers. Every headline benchmark is replayable against public data — or it is not a benchmark, it is marketing.
Mirror Fractal is, for now, deliberately a one-person operation. The only way to hold a single idea steady across many domains is to keep the organisation small enough that the idea doesn't blur.
We collaborate with other researchers, with engineering partners, and with end users who are willing to try things that aren't yet on anyone's roadmap. We don't scale headcount. We scale what we know.
If you'd like to work on a problem together — industrial partners, PhD candidates looking for a weird co-author, or anyone with neuromorphic data they can't compress — the door is open.
Write to the labEvery claim in our work is backed by a public deposit — preprint, wheels, checksum, demo. If it isn't on the timeline, it didn't happen yet.
Lossless compression for neuromorphic event-camera streams. Python wheels for Linux / macOS arm64 / Windows. WebAssembly build powering the browser demo. Preprint and manuscript archived with DOI.
pip install mfc-codec. Pre-built wheels for CPython 3.11,
evaluation and academic-use license. The surface API is tiny on
purpose: encode(events) → bytes, decode(bytes) → events.
WASM build running in the browser, no install needed. Drop a
.npy event file and watch the bit-exact round-trip and
the resulting compression ratio in real time.
US Provisional No. 64/034,974. Covers the method underlying the codec. PCT filing will follow within the 12-month priority window. Commercial licensing available — contact the lab.
Spatial container format for layered geometric data. Aiming for a public preview and a browser-native viewer in Q2. Preprint to follow.
Internal ablations underway on a small-model test bench. First external benchmarks expected mid-year.
We publish in places that don't disappear. DOIs at CERN-hosted Zenodo. Python wheels on PyPI. Public mirrors on GitHub. Preprints — arXiv when the numbers are ready and the patent filings permit.
Read the preprintCollaborations, licensing, residencies, technical questions, anything that sounds like a real person — all of it lands in the same inbox. We usually reply within a few days.