Visual designer living on the sunny side of the planet — building the tools I want to design with.
tsevis.com

I am a visual designer, and almost everything in this account is a tool I needed and could not buy. Photographs into instrument imagery, maps into editable vector artwork, a crowded sheet of icons into a named and separated library, colour theory into something you can actually move a slider on.

They are mostly Python desktop applications: local-first, offline by default, no cloud endpoints, and built for a real production workflow rather than a demo. Several are named after the Greeks who got there first — Hipparchus, Heron, Diophantus — which is not a theme so much as a habit.

_{Hipparchus — Santorini, fetched live from OpenStreetMap through Overpass and rendered straight out to layered, editable SVG.}

My interest in computer vision is a designer's interest: I want a machine to understand a picture well enough that I can restructure it. Not to classify it, not to caption it — to take it apart along the seams that actually matter and hand me the pieces.

That splits into two halves I keep using together.

Learned vision, run locally. Open-vocabulary detection and segmentation (GroundingDINO, SAM 2.1, VitMatte) do the semantic layering in Skiagrafia; a local vision-language model (Qwen2.5-VL / Qwen3-VL, through either Ollama or llama.cpp) gives Shapearator its semantic naming, so an extracted icon lands as heart.png rather than icon_003.png; monocular depth, subject matting, and SAM 3 region materials give Heron the scene understanding its physics needs. The constant is that inference is local and deterministic — model loaders are pinned to local_files_only, and an engine that reaches for the network mid-render is neither offline nor reproducible.

Classical, geometric computer vision, which is still where most of the work happens: centroidal Voronoi cells, Delaunay meshes, hex grids and circle packing as derived map geometry in Hipparchus; shape detection, cropping and canvas normalization in Shapearator; and in Diastasis the overlap graphs and colouring search that separate tangled vector artwork into production plates — and can prove the layer count is minimal.

And underneath both, colour as a measurable thing: perceptual spaces (Oklab, CIELAB) for ramps that climb monotonically, and Metelli's equations for perceptual transparency in Garau Armonie.

The through-line is simple enough: I want tools that find the structure in a picture, not tools that summarize it.

Hipparchus Desktop vector cartography. Live OpenStreetMap through Overpass, local .osm.pbf, vector tiles, Natural Earth, Overture, terrain DEM and VIIRS night lights — out to layered, editable SVG that opens clean in Illustrator. A shelf of cartographic presets, plus derived Voronoi, Delaunay, hex-grid and circle-packing geometry.	Heron Photographs re-synthesized as if captured by an instrument that sees past human vision — thermograph, fluoroscope, intensifier tube, NIR, Kirlian plate, schlieren bench, lumen chamber, spectroscope. Eight physically distinct engines on one pipeline. A physics simulation, not a gradient map. Artistic simulation — never a measurement or a diagnostic.
Garau Armonie A studio to teach Augusto Garau's colour theory and simulate his perceptual Colour Transparency, in memory of an unforgettable professor. One framework-agnostic TypeScript engine drives both a web studio and a Photoshop plugin, so a colour reads identically wherever you meet it. Live demo →	Skiagrafia Semantic vectorizing and masking. Local ML segments an image into objects and their parts, then writes both vector (SVG) and bitmap (TIFF/PNG) layers. 100% local inference, with a first-run wizard that fetches the weights so a fresh clone simply works.
Shapearator One crowded sheet of shapes in, a tidy library out — each icon detected, cropped, normalized onto a shared canvas, and exported to PNG/JPG/TIFF/SVG with per-icon metadata. Ollama and llama.cpp are interchangeable first-class backends behind a single vision-client contract.	Diophantine An encryption and compression companion, named for Diophantus of Alexandria. The one thing here that has nothing to do with pictures.

Python · NumPy · SciPy · OpenCV · shapely · scikit-learn · PyTorch · Tkinter · skia-python · FastAPI · Streamlit
TypeScript · React · Vite · Adobe UXP
Local inference · Ollama · llama.cpp · SAM · GroundingDINO · Whisper
Output that matters · SVG · GeoTIFF · PMTiles · TIFF

tsevis.com · all repositories