Inputs.
Body measurements collected at reservation. Fit history — pieces kept, returned, altered. Post-wear feedback when offered. Each input is voluntary. Each is kept locally per wearer.
THE STUDIO
Where the engines run.
Two models sit upstream of every piece. Here is how they work.
FIT INTELLIGENCE
The fit engine.
A model that predicts how a pattern will sit on a specific body, before the piece is sewn for that wearer.
What the model does.
Given a new pattern (sleeve length, shoulder slope, ease at chest, drop), the model predicts the size that fits closest and suggests a single seam adjustment that would make it sit cleanly. The output is a per-wearer pattern instruction, not a generic size letter.
What it is, today.
The fit engine is in early calibration. It runs on a small but growing dataset of reservations and fit history. Compute is cloud-based; we are evaluating NVIDIA preferred-pricing partners for sustained training cadence as the wardrobe grows.
GENERATIVE DESIGN
The design engine.
A loop that lets the studio explore silhouette, drape, and proportion in software, before a single sample is cut.
How the studio uses it.
The designer sketches an intent. Image and 3D models surface variations on silhouette, drape, and seam placement. The studio chooses a shortlist by hand. Sampling begins from that shortlist, not from a blank page.
Why it changes the cycle.
A studio without the engine commits to a sample early. A studio with the engine commits to a sample late — after a thousand variations have already been seen and rejected. Less waste of fabric, less waste of time, more conviction in the chosen cut.
What it is, today.
We use diffusion models for early silhouette exploration and 3D drape preview tools for proportion checking. Both run on GPU cloud. The output of this engine is the pattern that the fit engine then adjusts per wearer.
THE LOOP
Why both engines, together.
Either engine alone is interesting. Together they make the wardrobe a closed loop.
The design engine produces a pattern. The fit engine maps that pattern to the wearer. The piece ships. Wear data — what was kept, what was returned, what was altered — returns to both engines.
The next drop’s design engine knows which silhouettes were worn and which were returned. The next drop’s fit engine knows the wearer’s body better than the last. Each drop is sharper than the one before.
This is the only architecture that makes a small brand defensible against scale: every wearer increases the precision of every piece. The wardrobe compounds.
RESEARCH PREVIEW
Two instruments.
Two instruments for thinking about a garment before it exists. The first measures. The second speculates. Both run as simulations.
ARCHITECTURE
What’s underneath.
What this is built toward. Some of it runs today; some of it is in development. Each is named so you can verify it.
01 / 04
Body geometry inference
Body landmarks resolved from a single front-view image.
- 33-keypoint topology
- Browser-side inference
- Front-view + side-view modes
MediaPipe BlazePose · ONNX Runtime
02 / 04
Fit prediction
A small network maps measurements and landmarks to per-pattern adjustments.
- Per-region confidence scoring
- Pattern adjustment deltas
- Fit history feedback loop
PyTorch · ONNX Runtime
03 / 04
Generative exploration
Conditional silhouette synthesis from a curated couture archive. (Planned.)
- Drape and proportion conditioning
- Studio shortlist surfacing
- LoRA fine-tune over couture archive
Stable Diffusion XL · LoRA fine-tune
04 / 04
Inference acceleration
Compiled inference graph targeting NVIDIA accelerators.
- TensorRT graph compilation
- CUDA 12 runtime
- H100 cloud target
NVIDIA TensorRT · CUDA 12
TRANSPARENCY
What we will and will not do.
We will.
- · Show the wearer why a size was suggested.
- · Let the wearer override every prediction.
- · Use returns as signal — never as a customer cost.
- · Keep the loop small enough to actually close.
We will not.
- · Train on a wearer’s data without consent.
- · Surface trends back to the wearer (“people like you wore…”).
- · Sell signal to other brands.
- · Pretend the model is finished. It isn’t, and won’t be.
The intelligence is in the cloth, not the campaign.
engines, drop 01