Inside Manifold
This page explains how Manifold's model uses its taxonomy and how it is served. It then covers the API platform and where the data lives.
The model and the taxonomy
Manifold labels each photo with Qwen3.5-9B, an open vision-language model. A model that reads the photo and the list of labels together can follow a taxonomy without being retrained on it.
The taxonomy has trade and phase on top and defect types beneath. The model works down it one level at a time. It picks the trade and phase first, then a defect type within it, with one more step in between for the largest groups. At each step its answer is constrained to the labels allowed there, and it decodes greedily, so the same photo gets the same label.
Confidence is read from the model’s own token probabilities, with no prompt asking the model how sure it is. A low-confidence answer still comes back, with a flag, so the customer decides what to do with it.
Serving
The model runs on AWS SageMaker, on one NVIDIA A10G GPU in the UAE region. It scales to zero. Once 10 minutes pass without a call, the GPU is released, so idle time costs no GPU hours.
The first call after a quiet spell gets a 503 that asks the caller to retry in 5 minutes, while the GPU starts and the model loads, which takes several minutes. After that a photo takes under 2 seconds (median).
The API platform
Customers call the API with a key. Only a SHA-256 hash of each key is stored, with a short prefix to find it, and keys are compared in constant time, so a copy of the database gives away no usable key.
Each team has a rate limit over a sliding one-minute window, kept in Redis, and a monthly image quota. Plans are monthly and sized by image volume. Usage is metered per call and billed through Stripe. Sign-in to the dashboard goes through Amazon Cognito.
Where the data lives
Everything that touches a photo runs in AWS’s UAE region. That means the model, the API, the database and the cache. Storage is encrypted, and the database sits in private subnets. Photos are not stored, and only their size is recorded.
The website is served from outside the region, but photos go from the browser straight to the API. The only data that leaves the region is billing metadata, meaning the customer’s email and team name, which go to Stripe.