Re-coding BeReal

The front-end is a mobile-first PWA, and its defining moment is capture. When the prompt fires, both the front and rear cameras open at once: the rear shows what you are doing, the front shows your reaction, and the post carries your location. There are no filters and no retakes, because the constraint is the whole point of BeReal.

I built that capture screen and the submission flow behind it: opening both camera streams, composing the two frames into one post, and handing it to the API. The interesting work is in the constraints, getting two simultaneous camera streams to behave on a phone, and refusing the affordances (filters, retakes) that every instinct says to add.

On the back-end the photo arrives as an upload and has to be stored somewhere fast and cheap. I built the pipeline around Multer's in-memory storage, so an uploaded image is held as a buffer rather than written to the server's disk, and streamed straight to Cloudinary with an upload stream.

Keeping the file in memory and streaming it out means no temporary files to clean up, no disk to fill, and a server that stays stateless. The endpoint accepts the two camera images together (front and back), so the dual photo is a first-class shape in the API, not two unrelated uploads.

Storage is also where the images get shaped, and I let the CDN do the heavy lifting. The main photo is auto-cropped with content-aware gravity to a consistent height, and profile pictures use face-aware cropping (Cloudinary's face gravity) into square thumbnails, so a portrait stays centered on the face instead of the corner of a room.

Offloading the transforms to Cloudinary kept the API thin and the results consistent: format conversion, sizing, and smart cropping all happen on delivery, driven by a few transformation options rather than image-processing code I would have had to maintain.

Because the pipeline is the product, I covered it with integration tests rather than trusting it by hand. A Jest and Supertest suite hits the real endpoints (authentication, publications, and the rest) and asserts on the responses, so a change that breaks upload or feed visibility fails loudly instead of silently.

Writing those tests is also what made me a more careful reviewer on the team. Once the API had a safety net, code review could focus on design rather than on catching regressions by eye.