Care labels, certificates, delivery notes, and now the garment itself. TexFile reads the evidence at line speed and turns it into verified item records with calibrated confidence and full provenance. Every claim, traced to its evidence.
Germany collects around 1.3 million tonnes of used textiles a year, most of it sorted by hand into hundreds of fractions. At line speed nobody can reliably say what a garment is made of, so good material gets downgraded and less than one percent becomes new clothing.
Illustrative artifacts
Care label
62% CO · 38% PET
Unverified
Composition asserted on a worn label. No evidence trail.
Test certificate
Cert Nº ····· · scope ·····
Unverified
Issuer and scope cannot be traced to a source.
Delivery note
Lot ······ · 1,240 kg
Unverified
Quantities never reconciled against goods received.
Recycled content claim
Recycled content · 38%
Unverified
A claim no measured record supports.
Supplier declaration
Self declared · unsigned
Unverified
Declared by the supplier, countersigned by no one.
Origin statement
Origin · EU · undisclosed
Unverified
An origin no upstream document can place.
Chain of custody
Custody ····· → ·····
Unverified
A chain with links that were never recorded.
S2Regulation clock
The dates are fixed. EU rules now demand item level data on what a garment is made of, and penalties for unproven claims reach four percent of annual turnover.
25
Days
:
07
Hrs
:
36
Min
:
27
Sec
Unsubstantiated green claims banned
Environmental claims must be substantiated, with penalties of up to four percent of annual turnover.
TexFile is built by three founders with equal equity, gender mixed at executive level, spun out of a Horizon Europe research project at Constructor University Bremen. The same three people built and validated the system TexFile deploys. No external hiring is required to execute the plan.
DE
Danial Eyvazi
CEO · Product
Built the vision language capture pipeline and its orchestration in an EU research project at Constructor University Bremen.
NM
Nasim Maleki
COO · Commercial & Data
M.Sc. Data Science for Society and Business, Constructor University Bremen. Executive leadership with equal equity is a standing gender balance commitment.
AI
Amirali Iranmanesh
CTO · Engineering
Carries the validated pipeline onto production hardware at the sorting facility.
Mentor and coordination · Prof. Dr. Omid Fatahi Valilai — Mentor · Coordinator
The idea began on a sorting line, inside an EU research project at Constructor University Bremen. We saw it from the inside. At line speed nobody can say what a garment is made of, so the data the new rules assume is never recorded.
Origin
EU Research Project
Horizon Europe
Constructor University Bremen
02The scale of the loss
Germany collects about 1.3 million tonnes of used textiles a year, most of it still sorted by hand. Barely any of it becomes new clothing again. Good material is downgraded because no one can prove what it is.
Where it all goes
Reused62%
Downcycled or burned37%
Recycled into new clothingless than 1%
bvse Alttextilstudie · Ellen MacArthur Foundation
03The idea
Almost every garment already carries its own documentation, sewn in at the seam. We read that label at line speed and turn messy text into verified facts, each one with a calibrated confidence.
Fragments to record
04A credit bureau for material claims
Each item becomes a knowledge graph built to the EU Digital Product Passport model and GS1 standards. Unreliable fragments in, a record the whole value chain can rely on out. One engine serves several customers.
At line speed
23,800 items per hour
50 ms routing call
05What remains
The material returns to the loop, routed by evidence instead of guesswork. Its verified record joins a databank that keeps nature green and safe, and proves every claim when the auditor asks.
Confidence gate
25%route automatically at 0.9 confidence
The rest go to assisted review, and every correction is written back with its provenance.
04TEXFILE · THE IDEA
The fate of a garment is decided in two seconds of judgement, and nothing of that decision survives as data.Yet almost every garment already carries its documentation, sewn in at the seam.We read that label at line speed and turn it into a verified record that routes material to reuse and recycling.
A multi view capture station photographs each garment at the line. A vision language pipeline reads composition, care, origin and brand from labels in any language or condition, normalizes them into controlled EU vocabularies with calibrated confidence, and assembles a per item record aligned with the EU Digital Product Passport model, GS1 EPCIS and PROV-O. Sorters route more garments into reuse and fiber to fiber recycling. Brands and producer responsibility schemes buy verified chain of custody evidence for stock that carries no digital identifier.
Measured, not asserted
23,800items per hour sustained at 16 concurrent requests
340label images in the frozen test set, hash locked and never used for training
25%of items route automatically at the 0.9 threshold, with no observed errors
50 msceiling for a routing answer from the record API
Against what exists
Manual sorting
Sorter judgement in seconds. No data output.
TexFile keeps a persistent, auditable record of the decision and its evidence.
NIR sorting hardware
The dominant surface fiber, as a material signal.
TexFile adds declared blends, care, origin and brand, and emits a record instead of a signal. It complements NIR where installed.
Generic document AI
Key value extraction built for flat office paperwork.
TexFile brings a waste stream label corpus, textile vocabularies, per field calibration and DPP aligned output.
Brand side DPP platforms
Distribution of data brands already hold on new products.
TexFile originates data at the point of sorting for untagged stock and feeds these platforms through open standards.
Origin
TexFile is a spin off of a Horizon Europe research project at Constructor University Bremen. The pipeline was validated against real waste stream material and line side capture before the company work began.
The moat
The defensible core is not the foundation model, which anyone can license. It is the waste stream label corpus, corrected through every assisted review, and the compounding per item knowledge graph built on it.
Next
Deployment on edge units at two operational sorting facilities, a nine month live pilot, an independent audit of calibration and provenance, and conversion of both pilots into the first commercial agreements.
Multi view, calibrated capture. A quality gate scores each image and says which view to retake.
02Extraction
Vision models read worn labels into structured facts, each citing the verbatim label text.
03Reconciliation
Conflicting sources resolve by a transparent weighted vote. Verified means corroborated.
04Confidence & provenance
Calibrated per field: 0.9 means nine in ten. Each fact auto commits or goes to human review.
05Record assembly
A knowledge graph per item, aligned with the EU DPP model and validated against SHACL shapes.
06Record store & API
Immutable snapshots, version history, a read only API. Routing fast enough for a live line.
Item record
REC-0047
Illustrative data
Fibre composition
62% CO / 38% PET
92%
Origin
EU · undisclosed
71%
Recycled content
38%
54%
DPP Ready
06TEXFILE · PROOF
Every field traces to its evidence.
A TexFile record is not an assertion. It is a chain of evidence: every value carries a calibrated confidence and a link back to the document it was read from.
Item record
REC-0047
Illustrative data
Select a field to trace it to its source.
SRC · care-label · p.1
SRC · care-label · p.1
SRC · care-label · p.1
SRC · material-cert · §2
SRC · material-cert · §2
SRC · delivery-note · L4
SRC · delivery-note · L4
SRC · supplier-decl · §1
DPP Ready
↳Source evidence
Care label scanSRC · care-label · p.1
Material certificateSRC · material-cert · §2
Delivery noteSRC · delivery-note · L4
Supplier declarationSRC · supplier-decl · §1
↳How a record is builtDocuments in. A scored record out, with provenance on every field: queryable, exportable, DPP ready.
↳What every record proves
A stable identifier for every item.
Provenance from every field to its source evidence.
A calibrated confidence on every extracted claim.
Shape validated before the record is sealed.
Exportable to the Digital Product Passport format.
TexFile reads the garment itself, not only its paperwork. Calibrated images become condition, category, measurement and brand signals, and every one of them flows through the same reconcile, score and provenance pipeline as a label fact. No visual claim enters the record without a source.
01Condition & damage
Holes, tears, stains, open seams and fraying, each with a severity. The condition grade that follows is deterministic and versioned: the same images always yield the same grade.
02Category & attributes
Garment type classified automatically, with attribute suggestions the operator can accept or correct.
03Measurements
Key dimensions in millimetres, taken from a calibrated flat lay image rather than a tape measure.
04Brand & label reading
Text read from labels and graphics, plus brand logo detection. Each fact cites the image it was read from.
An agentic layer turns a verified garment record into outputs a resale or upcycling partner can use the same day. It works under one strict rule, enforced in code rather than promised in prose.
Real resale listings are found and cited as comparables. The band and the sell through estimate are computed by versioned, auditable arithmetic, never made up by a model.
02Repair suggestions
A deterministic repairability check, with guidance drawn from a verified repair library.
03Defect fix preview
AI imagery shows the garment with a defect repaired, quality gated so the model provably touched only the marked region.
04Hero imagery
Photorealistic images of the garment worn by AI models from a diverse, versioned persona roster.
05Grounded listing copy
An AI written title and description where every hard fact, from grade and fibre to measurements and price, is projected from the verified record through an allowlist.
Market grounded price bandIllustrative data
P25MEDIANP75
Computed from cited comparables. Never estimated by a model.
Sell through · estimated
Defect fix previewIllustrative data
AI GENERATED · ART. 50
Pending review
Approved · named reviewer
Exported · EPCIS custody event
↳The guardrails
Disclosed
Every generated image carries a visible AI disclosure band, burned into the pixels. EU AI Act Article 50, by construction.
Human approved
Every AI asset starts as pending review. A named person approves, rejects or revokes it, on an append only log. Partners only ever see approved, disclosed media.
Custody tracked
Reuse routed garments enter a partner queue matched by capability: resell, upcycle or repair. Every transition is logged, and an export mints an EPCIS custody transfer event.
Byte stable packs
A partner export is a byte stable bundle, listing data, approved disclosed images and the item code, projected from the verified record.
09TEXFILE · SURFACES
One engine. Seven surfaces.
Everything the engine verifies is worked, reviewed and audited somewhere. Seven surfaces cover the path from the sorting line to the auditor's desk, each built for the hands and the light it is actually used in.
01
Operator kiosk
A dark, touch first console at the line: capture a label, get a routing answer at arm's length. Works offline and syncs when the connection returns.
02
Facility workspace
The shift's dashboard: items processed, auto route rate, verified coverage, what is held for review, and a live activity stream.
03
Assisted review
A keyboard first queue for low confidence facts: confirm or correct, grouped by field. Every correction is written back with its provenance.
04
Item record viewer
Per field confidence and sources, the conformance report, the EPCIS timeline, the GS1 Digital Link code and an interactive knowledge graph view.
05
Verification & audit
A batch's recycled content statement and chain of custody with a mass flow diagram, exported deterministically for the auditor.
06
Admin panel
Facility registration, license management, the approval console for generated media, and oversight of every partner handoff.
07
Partner portal
A live offer queue and a composed garment dossier, imagery, condition, repair advice and the price band, through to export.
Installable as a PWA. Capture continues through connectivity loss and syncs when back online.
Real time
Live event streaming, degrading gracefully to polling.
Responsive
Phone to wall mounted display, light and dark themes, density preferences.
10TEXFILE · INTEGRATION & TRUST
Built to integrate. Built to be audited.
TexFile sits upstream of DPP platforms and beside the systems a facility already runs, so the edges are part of the product: a typed API for partners, and a security posture published as capabilities rather than promises.
↳The API
Typed REST
Generated from an OpenAPI 3.1 schema, with versioned JSON Schema contracts for every shape partners integrate against.
Records, read only
Query by item, version, batch or facility. Immutable snapshots, full version history, JSON-LD and RDF export.
GS1 Digital Link
Every item carries a GS1 Digital Link URI and a scannable code, served by a conformant resolver.
Report endpoints
Recycled content and chain of custody reports, every figure traceable to fact level provenance, coverage openly disclosed.
Partner API
Token auth, the offer queue, the garment dossier and listing pack export, with live events and a polling fallback.
↳Security & trust
Authentication
License key and admin authentication with short lived tokens, session revocation and refresh token rotation.
Tenancy
Role based access, scoped strictly by facility and partner. A partner only ever sees what was offered to them.
Audit trails
Append only: logins, admin actions, handoff transitions and every media approval decision.
Encryption
Personal data encrypted at rest, license keys stored only as hashes, modern password hashing.
EU residency
EU region infrastructure with data residency guardrails, an on site edge console for the facility floor, and AI tracing that is EU pinned and payload blind.
Published process
Hardened web and API layers, with a published methodology: STRIDE threat modeling, OWASP ASVS L2 verification and automated scan gates in CI.
11Roadmap
From a research prototype to the trust network: what is shipped, what we are building, and what comes next. Every line traces to a dated record.
FoundationFoundation
Shipped2026 · 06technology
Verification prototype validated
Label reading and normalization into controlled vocabularies, validated against a frozen test set of 340 label images with calibrated confidence at the 0.9 operating point.
Shipped2026 · 07company
Pilot data access across a European consortium
Letters of Intent from consortium partners across collection, sorting, recycling, sensing and research secure pilot facilities and real waste stream imagery.
[FOUNDER: confirm LOI dates before publishing specifics]
Verification EngineVerification Engine
Buildingproduct
Pilot installation at a partner sorting facility
An industrial scan station on a running line: camera and edge integration, hardened to sustain 18,000 items per hour.
Buildingtechnology
Robustness on damaged and multilingual labels
Accuracy hardening on the worst material first: worn, damaged and multilingual labels from real waste streams.
Shipped2026 · 08product
DPP aligned record export
Standards conformant passports from the record store: JSON-LD export, GS1 Digital Link resolution, SHACL validation before anything ships.
Nextproduct
Verification service v1
Verified recycled content and chain of custody records for brands and EPR schemes, with a product carbon footprint from the same evidence.
Garment VisionGarment Vision
Shipped2026 · 08technology
Garment vision joins the record
Reading the garment itself, not only its paperwork: condition and damage with a deterministic grade, category and attributes, calibrated measurements and brand signals fold into the item record as added evidence sources.
Shipped2026 · 08product
Upcycling & resale layer
A verified record becomes resale ready outputs: a price band computed from cited comparables, repair guidance from a verified library, disclosed AI imagery behind human approval, and a partner handoff that mints EPCIS custody events.
NOW
The Trust NetworkThe Trust Network
Nextcompany
Incorporation and first commercial installation
The pilot converts into the first commercial installation, and two more facilities follow within twelve months.
[FOUNDER: confirm timing after the funding decision]
Visioncompany
The trust network
One shared verification layer across sorters, recyclers, brands, EPR schemes and cities. Each installation improves recognition for everyone already live.
12Partners
Built for everyone who moves material.
TexFile is a shared record. Wherever a garment is collected, sorted, recycled, sold or certified, the same evidence travels with it.
