kapra

A garment forensics instrument. Photograph a piece of cloth; kapra reconstructs its genome — weave, colour, texture, weight, finish — using real signal processing, and tells you exactly how far each reading can be trusted.

Not an AI wrapper. There is no model guessing on your behalf: a 2D Fourier transform measures the weave, k-means in CIELAB measures the dye, Local Binary Patterns measure the texture. Every claim carries the measurement that produced it, and the instrument says plainly where its knowledge stops.

Zero model weights. The core runs on NumPy and OpenCV, locally. Clone it, run one command, and it works — offline, with no API keys.

---

What it tells you

A report opens in plain English, then shows the measurements behind it:

A midweight indigo twill weave.

Fabric Twill weave	Colour indigo #33466b	Weight Midweight
Threads step across to form a diagonal. Think denim, chino.	Dyeing is very even (91% consistency).	~200–320 g/m² — an estimate; a photo has no scale.

Below that sit the real figures: the Fourier power spectrum, Gabor orientation energies, the CIELAB palette, LBP histograms, confidence meters and the evidence list for every gene.

The honesty system

Every claim wears a tier, and the schema refuses to serialise one without evidence — so an unsupported claim cannot reach the page.

Tier	Meaning	Example
MEASURED	Computed directly from pixels. Deterministic, reproducible.	Thread repeat via FFT, CIELAB palette, LBP texture
ESTIMATED	A prediction carrying confidence and, where meaningful, an interval.	Weave family, weight class (g/m² range)
INFERRED	A hypothesis, labelled as one, citing what led to it.	Construction & process (planned)

The split is deliberate. The thread repeat is read to within ~1% of truth, so it is MEASURED. Naming the interlacing is a judgement, so it is ESTIMATED. GSM cannot be known from a photograph at all, so it is reported as a class with a wide interval and low confidence — never as a fabricated number.

Genes

Gene	Tier	Method
Weave Structure	MEASURED	2D FFT · GLCM · Gabor bank
Weave Family	ESTIMATED	Off-axis lattice ratio (√2 plain, √5 satin, 2√2 twill)
Colour & Dye Evenness	MEASURED	CIELAB k-means · CIEDE2000
Texture Fineness	MEASURED	Local Binary Patterns · Laplacian energy
Finish & Sheen	MEASURED	Specular-highlight distribution
Weight Class	ESTIMATED	Thread density from the measured repeat

Validated results are in docs/BENCHMARKS.md.

---

Quick start

Requires Node 20+, pnpm, and Python 3.12+.

pnpm install
pnpm run bootstrap     # builds the schema, creates the venv, installs the CV stack
pnpm run dev           # engine :8000 + web :3000, one terminal

Open http://localhost:3000 and sequence one of the six demo swatches — no photo of your own required.

pnpm run dev runs both servers with prefixed output; Ctrl+C stops both.

Layout

apps/
  web/      Next.js front end — editorial multi-page site + the forensic report
  engine/   FastAPI CV service — the extractors, orchestrator and persistence
packages/
  genome-schema/  the versioned Genome schema: ONE source of truth (TS + Python)
  ui-tokens/      design tokens: bone paper, ink, cyanotype bands, oxblood accent
docs/       plan, architecture, data/licences, benchmarks, genome spec, deployment
scripts/    dev.mjs (run everything) · bootstrap.mjs (set up a fresh clone)

Architecture in one paragraph

packages/genome-schema holds the canonical JSON Schema. TypeScript types are generated from it; the Python Pydantic models mirror it, and a conformance test fails if they drift. The engine exposes extractors as plugins — adding a gene is adding one file that registers itself; the orchestrator discovers it, runs it, times it, and isolates its failures into warnings. See docs/ARCHITECTURE.md.

The demo swatches

The six specimens in apps/web/public/swatches/ are generated, not stock photos — real interlacing patterns rendered with yarn shading, slub and fibre fuzz by apps/engine/scripts/generate_swatches.py. No licensed assets, and the instrument reads them exactly as it reads a real macro.

Regenerate them with:

cd apps/engine && .venv/Scripts/python scripts/generate_swatches.py

Quality

pnpm run typecheck && pnpm run lint && pnpm run test && pnpm run build
cd apps/engine && .venv/Scripts/python -m pytest -q

TypeScript strict (with exactOptionalPropertyTypes), ruff, mypy --strict, pytest, vitest, and GitHub Actions across both toolchains.

Deploying

Web on Vercel; the Python engine on any container host (Render has a real free tier and a render.yaml blueprint is included). See docs/DEPLOYMENT.md.

Status & limits

Six genes live across two tiers, end to end: upload → analysis → stored genome → report → archive. What it deliberately cannot do — fibre content, true GSM, brand or authenticity — is documented on /method and in docs/DATA.md.

Research use. Dataset licences for the planned ESTIMATED classifiers are documented rather than glossed over.