AI agent for fashion · AMD Hackathon ACT IIIAI agent for fashion brands · AMD Developer Hackathon ACT III
Make only what’s already sold.
- Customer voice
- Creative direction
- Supplier quotes
- Pre-orders
- Forecast
Illustration - the gauge is a scale, not a reading. Every number in the studio below is measured or shown as “-”.
What it does
Atelier reads a trend from a mood board, sketches the garments and takes your direction in plain words. Its agents write the tech pack, brief the mills and open a pre-order - production is ordered only once pre-orders cover it.
System status · live from /api/health
Checking services…
01The problem
Fashion bets months ahead, then lives with the leftovers.
The industry commits to production on a forecast, long before demand is known. Atelier inverts the order: design fast, sell first, make what has sold.
- i
The lag
A trend is spotted on a runway or a feed. The garment reaches the shelf after sketches, tech packs, mill emails and sample rounds - by then the moment may have moved on.
- ii
The guess
Buy quantities are committed before a single customer has said yes. Too few and the line sells out; too many and it ends in markdowns, or worse.
- iii
The paperwork
Taste is the scarce part, yet most hours go to specs, costings, RFQs and supplier comparisons: structured, repetitive work that is easy to get wrong.
- 2.5–5bn items
- surplus garments produced in 2023, worth $70–140 bn in sales
- 4–9%
- of textiles on the EU market are destroyed before they are ever used
- 19 Jul 2026
- large companies may no longer destroy unsold apparel and footwear in the EU (ESPR, Art. 25)
02How it works
Four chapters. One creative director.
The studio below runs as four chapters on one page. In each, you make a call and the agent does the work around it, leaving something concrete behind.
- I
The signal
Trend ingestion
You
Drops in a mood board: a preset, uploaded images or image links.
AtelierA vision-language model names the trend, reads its silhouettes, palette and fabrics, and scores commercial viability.
Leaves behind: A structured trend read
- II
Four concepts
Generative ideation
You
Asks for four concepts, or a small capsule collection.
AtelierWrites the image prompts from the trend and renders the garments on the diffusion server.
Leaves behind: Four seeded concept renders
- III
The vibe check
Human in the loop
You
Picks a concept, directs it in plain words (typed or spoken) and approves one look.
AtelierTurns the brief into prompt, palette, silhouette and construction changes, reads them back, then renders four directed variations.
Leaves behind: An approved design with its direction
- IV
Into production
Agent pipeline
You
Edits the tech pack if needed and can overrule the supplier pick.
AtelierFive agents run in sequence, from tech pack to purchase order, gated on real pre-order demand.
Leaves behind: Tech pack PDF, RFQs, extracted quotes, a live pre-order page, a forecast, a purchase order
03The agents
After approval, five agents take over.
They run in sequence on the approved design, each handing a typed artifact to the next. Quotes are read from real documents, pre-orders are real Shopify orders, and the PO goes out through Evolus. If a service is missing, that step stops with a setup hint - nothing is made up.
- 01
Tech pack
Evolus agent · Checking…LLM · Checking…Takes
The approved design, its trend read and the creative direction
Hands on
Bill of materials, measurements, construction notes, colourway, costing and retail price, validated against a schema
Runs on · An Evolus agent on the AMD vLLM model, grounded in the brand book - or AMD vLLM directly
- 02
Supplier outreach
Evolus outreach · Checking…Takes
The tech pack and the target sourcing region
Hands on
Request-for-quote emails to the brand's own suppliers, each in the supplier's language (English kept alongside), with the exact colours to lab-dip against
Runs on · The agent drafts; an Evolus workflow sends them to the brand's suppliers (or a pilot inbox)
- 03
Supplier quotes
Evolus extractor · Checking…Takes
The mills' quote documents (PDF or image), uploaded or emailed back
Hands on
A scored shortlist (price 45%, MOQ fit 30%, lead time 25%; a criterion the quote doesn't state is left out, never guessed) and a recommendation
Guardrail hand-off. If the pick breaks a brand guardrail (margin floor, cost cap, lead time, MOQ, certificates), the agent stops and hands the call to you: approve, switch, or negotiate in the mill's language.
Runs on · An Evolus document extractor reads each document; deterministic scoring; the agent writes the rationale
- 04
Shopify pre-order
Shopify · Checking…Takes
The final render, title, copy and retail price
Hands on
A real product on a Shopify store, with a QR code to pre-order it
Runs on · Shopify Admin API, or an Evolus workflow
- 05
Pre-order demand → PO
Shopify · Checking…Evolus PO · Checking…Takes
Real Shopify orders and the threshold (run size or supplier MOQ, or a configured stage threshold)
Hands on
An explained forecast of how many to make; once the threshold clears, a numbered purchase order sent to the mill through Evolus (without Evolus it stays a draft)
Runs on · Atelier orchestrator; the forecast is closed-form statistics, no language model
Badges show each agent’s backing service right now, from /api/health. A failed step can be retried once its service is back.
04Human + AI
Direct it the way you’d brief a studio.
No sliders, no prompt engineering. You say what you want; Atelier turns it into concrete changes, reads them back, and carries them all the way into the tech pack.
You decide
- IWhich trend to chase
- IIWhich concept to develop
- IIIHow it should change
- IVWhether it goes to production
Atelier reads back
- Prompt
- What is added to and removed from the image prompt for the next renders.
- Palette
- The colourway before and after, when the direction re-colours the garment.
- Silhouette & fabric
- Changes to proportion and construction, and any fabric swaps (shell first).
- Into the tech pack
- Concrete hints carried into the tech pack the agent writes after approval.
The brain (an Evolus agent or AMD vLLM) interprets the instruction; the diffusion model renders four directed variations. If either is not connected, chapter III says so instead of showing a result.
Atelier heard
Your read-back appears here once you direct a concept.
Pick a concept in chapter III and say what you want - “darker”, “cropped”, “more utilitarian” - typed or spoken. What shows here is the model’s real answer; nothing is pre-recorded.
05Under the hood
Open models, an orchestrator and Evolus building blocks.
Open models served with vLLM and a diffusion server on AMD hardware; a FastAPI orchestrator that owns the business loop; Evolus for the brain and the outside world; a real Shopify store. Model names and the GPU below are what this instance reports right now.
The desk
Atelier studio
Next.js · React
Mood boards, concepts, direction, approval, the tech pack editor, quote documents, the pre-order tracker and the forecast.
The loop
Orchestrator
FastAPI · Python 3.12
Owns the typed contract and the five-step pipeline, records which real path produced each result, scores quotes, runs the pre-order gate, forecasts the run, renders the tech pack PDF and measures every inference call.
AMD models
GPU: not reported
- Checking…
LLM
Tech pack, direction, RFQs, supplier rationale
- Checking…
Vision-language model
Reads the trend from the mood board
- Checking…
Diffusion model
Concept and variation renders
Evolus
Driven through the Evolus API
- Checking…
Agent · the brain
Runs on the AMD model, grounded in a knowledge library (brand book, supplier catalogue)
- Checking…
Document extractor
Reads the mills' quote documents, field by field
- Checking…
Outreach workflow
Sends RFQs and counter-offers; email quotes come back by webhook
- Checking…
PO workflow
Sends the purchase order to the mill
Shopify
Admin API · dev store
- Checking…
Pre-order store
The real product page and the orders that unlock production
Honest by default
A step whose service is not connected stops with the error and a setup hint, and can be retried. Quotes, orders and renders are never invented; a local test double is labelled FAKE in red.
Measured
Every model call records model, host, GPU, latency and tokens per second; the Inference panel in the header shows them.
One contract
Studio and orchestrator share a typed API contract; every result records which real path produced it.
Model names, GPU and badges show what this instance is actually calling, from /api/health; a service that is not connected names no model.
06Business value
Less guessing, less dead stock.
For brands that sell direct and want to test more ideas without carrying more inventory.
Sell first, then make
Every approved design goes live as a real pre-order. The purchase order is drafted only once orders clear the threshold, and every order sharpens an explained forecast of how many to make.
Margin before a cut
The tech pack arrives costed. Swap a material and unit cost, margin and break-even update on the spot; save it and open quotes are re-scored against your guardrails - the mills’ prices stay as quoted.
Mills compete
Your suppliers get the same brief at once. Their quote documents are read field by field and scored on price, MOQ fit and lead time; the recommendation shows its reasoning, and you can overrule it.
Taste stays human
You make the four calls that matter: the trend, the concept, the direction and the approval. The agents do the rest.
Where the time goes · design & development
11 to 24 weeks before a style is ready to make. Atelier’s own figure: to be measured.
No pointer on the tape: Atelier’s time from approval to tech pack and RFQs has not been measured on real runs yet, so no figure is claimed. Sampling, production and shipping still take as long as the mill needs.
Selling first already works
>90%
of Stòffa’s revenue is paid before production. Prabal Gurung and Antonio Berardi take 20–25% of revenue as pre-order or made-to-order.
The studio is below · chapters I–IV
Take the chair. Pick a mood board.
Everything from here on is the working product, running against the backend shown in the header.
IChapter I · Trend ingestion
The signal
Customer voice · reviews, returns, tickets
Start from what your customers say
Drop in your own review, return or support-ticket export. Atelier counts what customers keep raising, quotes them word for word, and turns the patterns into tech pack changes you can apply to the design.
Files stay in this session’s memory on the Atelier backend. Email, phone and name columns are ignored.
No data to hand? Use a sample