Try live demo → ProductFor ShoppersFor BrandsFabric IntelligenceTechnology About Join Shopper Beta Request Pilot
Technology

Orchestrated intelligence. Not a wrapper.

Atelia is a multi-layer reasoning system. Deterministic where facts matter, like stock, price, and composition. Probabilistic where judgement matters, like style, feel, and fit.

The architecture

Six reasoning modules, one verdict.

Module 01

Intent Parser.

Converts messy human language into structured occasion, tone, and constraint signals.

Module 02

Preference Memory.

Builds a quiet, evolving model of how each shopper actually shops, without intrusive profiling.

Module 03

Catalogue Intelligence.

Enriches raw retailer data into AI-readable garment metadata across occasion, silhouette, and feel.

Module 04

Fabric Reasoning.

Translates fibre composition into lived experience. Drape, breathability, maintenance, durability.

Module 05

Tradeoff Engine.

Probabilistic comparison across competing options. Gained, sacrificed, priced.

Module 06

Explainability.

Produces the human-readable why this. Every recommendation is grounded in traceable evidence.

Hybrid intelligence

Deterministic where facts matter. Probabilistic where judgement matters.

Stock levels, prices, composition percentages, care labels. These are facts. They demand exact answers. Style, drape, quiet luxury, client-dinner appropriate. These are judgements. They demand probabilistic reasoning. Atelia blends both, and never confuses them.

Deterministic
  • Stock availability
  • Price and currency
  • Fibre composition
  • Care label text
  • Size availability
  • Returns policy
Probabilistic
  • Drape behaviour
  • Style adjacency
  • Occasion fit
  • Aesthetic tone
  • Confidence reasoning
  • Tradeoff weighting
Grounding principles

Six rules Atelia will not break.

01

Grounded only.

Atelia makes no claim it cannot trace back to verifiable garment, fabric, or price data. If it cannot be grounded, it is not said.

02

Explainable always.

Every verdict comes with the reasoning behind it. No black-box recommendations. The shopper sees how we got there.

03

Restrained, not loud.

We improve the shopper's ability to decide. We never decide for them. The tone is confidante, not authoritarian stylist.

04

Fashion-specific.

Built for garments, not generic ecommerce. Trained on the way fashion is actually evaluated by people who know it.

05

Privacy by restraint.

Preference memory is opt-in, low-resolution, and never sold. We do not need to know your address to recommend a coat.

06

Tradeoffs disclosed.

If a garment has a weakness, we say so. Atelia is not a sales tool. It is a confidence tool. Trust depends on honesty.

Data flow

A single query, through the engine.

Step 01, input

"Something that feels expensive but not flashy."

Raw natural language from the shopper. Ambiguous, mood-led, real.

Step 02, parse

Structured signals

  • tone, quiet
  • occasion, evening
  • silhouette, restrained
  • fabric, matte
Step 03, reason

Engine pass

  • catalogue match
  • fabric reasoning
  • tradeoff modeling
  • price-value check
Step 04, verdict

Explainable output

  • 6 ranked matches
  • scores, fabric, drape
  • why this reasoning
  • tradeoffs disclosed
The stack

Built on modern foundations.

A blend of retrieval-augmented generation, specialised fashion vision models, vector-based semantic search, and rule-based grounding. No single model. No single vendor lock-in.

Retrieval

Semantic Vector Search.

Multi-embedding catalogue index. Meaning similarity, not keyword. Updated continuously as catalogue evolves.

Generation

Reasoning LLMs.

Frontier models orchestrated through Atelia's reasoning layer. Never raw, always grounded.

Vision

Garment-specific vision.

Specialised models trained on silhouette, drape, finish, and fabric recognition. Built for fashion, not photos generally.

Rules

Deterministic layer.

Hard constraints on stock, price, size, care. Facts never enter the probabilistic layer.

Infrastructure

Streaming catalogue ingest.

Connects to Shopify, XML, custom feeds. Continuous enrichment as inventory changes.

Privacy

Restraint-first memory.

Preference memory is opt-in, low-resolution, never sold. Compliance-ready from day one.

Want to look under the hood.