How We Automated Catalog Image Extraction using Computer Vision & FastAPI

somyabhalani — Tue, 19 May 2026 19:24:39 +0000

For businesses in the stone, marble, and interior design industries, managing digital catalog assets is a massive headache.

When a new product catalog arrives as a 100-page PDF, design teams spend hours manually cropping out individual tile samples to upload to their websites or inventory sheets.

To automate this, we built Tile Extractor—a high-performance, automated parsing engine designed specifically to isolate tile samples from raw catalog documents.

How it Works (Under the Hood)

PDF Ingestion: The system uses a FastAPI backend to ingest multi-page PDFs. We process the pages using PyMuPDF to extract raw page vectors and high-res layout structures.
Object Detection & Border Cleaning: Instead of relying on slow, expensive cloud Vision APIs, we use local Pillow and OpenCV-based spatial algorithms. The engine analyzes:
- Edge density to isolate individual tile boundaries.
- Aspect ratios to filter out page noise (like page numbers or logos).
- Color distributions using RGB histograms.
Lossless Cropping: Once a tile is identified and classified, the engine performs a lossless crop directly from the PDF's high-resolution asset stream, ensuring no pixel resolution is lost.
Batch ZIP Packaging: The isolated tile PNGs are packaged into a single ZIP file and returned to the user instantly.

Why it Matters for B2B Automation

What used to take a human designer 2 hours now takes our engine 5 seconds. By running localized computer vision algorithms instead of cloud APIs, we eliminate usage fees and keep client data fully private.

If your business manages product catalogs, you can try the tool for free here:

👉 Try Tile Extractor: https://tile-extractor.onrender.com
👉 Explore our work: https://anantalabs.app/

How We Built a Contactless Digital Signature App inside the Browser (No Servers, 100% Private)

somyabhalani — Tue, 19 May 2026 19:21:46 +0000

Traditional digital signature platforms have two major issues: privacy and cost.

To sign a document, you have to upload sensitive agreements to a third-party server. And as a developer, running server-side document rendering and signatures can lead to heavy API bills and database management overhead.

At Ananta Labs, we wanted to see if we could build a completely secure, contactless alternative that runs entirely on the client side using browser-native AI.

Here is how we built AirSign.

The Architecture: 100% Client-Side

Instead of hosting heavy machine learning models on a GPU server, we compiled our hand-tracking models to run locally inside the user's browser.

Gesture Capture: We utilized MediaPipe's hand-landmarker models compiled into WebAssembly. This allows the browser to track 21 3D hand coordinates in real-time at 30 FPS using a standard webcam.
Contactless Canvas: Using WebGL, we map the index finger coordinate to a HTML5 canvas. We implemented a custom interpolation algorithm to smooth out hand jitter and render a fluid, realistic signature line.
Local PDF Stamp: Once the user finishes drawing their signature in the air, we generate the final document. The signature coordinate vector is parsed and stamped onto the PDF using a client-side library.

Why this is the Future of AI Integration

By moving the computation from the server to the client:

Absolute Privacy: 0 video frames, coordinate points, or document bytes are transmitted to any database.
Zero Server Overhead: The hosting cost for this app is exactly $0 since it runs on the user's CPU/GPU.
Instant Load Times: Zero network latency during signature interpolation.

Try it Yourself

AirSign is completely open and free to test. We’d love to hear your feedback on the hand-tracking latency and mobile performance:

👉 Try AirSign: https://airsign-red.vercel.app/
👉 Explore our work: https://anantalabs.app/