Aditya Rathi

Own product, 2026

Manifold, a paid API for construction-site photos

Manifold is my own product. It is a hosted API that sorts construction-site photos by trade, phase and defect type, sold on monthly plans by image volume.

Built with Qwen3.5-9B, AWS SageMaker, FastAPI, API keys, usage-based billing.

  • under 2 secondsmedian inference time per photo
    About this number

    Hosted classification API, median time for one construction-site photo. Measured on one ml.g5.xlarge SageMaker instance (NVIDIA A10G) with two to five requests in flight. As of 2026.

How a construction-site photo gets its labelA customer uploads a construction-site photo to an API with key authentication, metered per call. Qwen3.5-9B on AWS SageMaker in the UAE region, prompted with the construction taxonomy, labels the photo by trade and phase, with a defect type beneath. Numbers shown: under 2 seconds median inference time per photo.Construction-site photouploaded by the customerAPIkey authentication, metered per callQwen3.5-9B on AWS SageMakerthe construction taxonomy is encoded in the prompt< 2 s median inference time per photoTrade, phase and defect typefrom the construction taxonomy
Manifold's classification API, from upload to label.

Why it mattered

A property developer in Dubai was sorting construction-site photos by hand. Manifold sorts each photo by trade and phase, and by defect type beneath them, and the developer pays for it as an API.

In the product teaser, a site photo goes in and a category with its likely items comes out. The scores on screen are the model's confidence on one demo photo, so don't read them as accuracy. The video has no sound.

Approach

The construction taxonomy is encoded into the prompting of Qwen3.5-9B, served on AWS SageMaker in the UAE region. The API has key-based authentication, monthly plans by image volume metered per call, and published documentation.

Result

Inference takes under 2 seconds per photo.

What I got wrong

What I assumed going in
That latency would matter more than cost per call.
What I'd measure first next time
Accuracy for each trade and phase, on the customer's own photos.
What it can't do yet
Night and low-light site photos.