Executive Summary: Velo Logistics operates 10,000+ warehouse inventory terminals where truck drivers and operators log packages. Standard signature pads are prone to high physical wear-and-tear and require constant sanitization. Velo hired Ananta Labs to engineer a touchless, camera-based digital signature platform. The goal was to support contactless handwriting using only standard integrated webcams on low-spec client terminals, requiring 100% offline security with zero backend server overhead.
The Technical Challenge: MediaPipe vs Custom CNN
To capture accurate handwriting gestures, the front-end software layer must detect and segment hand coordinate patterns at 30 FPS. We analyzed two architectures:
1. Custom Convolutional Neural Network (CNN)
We initially built a custom CNN running inside TensorFlow.js. The model classified frames to determine the index finger tip coordinate. While highly specialized, it required extensive training datasets for varied hand angles, struggled with lighting conditions at loading docks, and consumed up to 85% of CPU threads on low-spec Celeron-based warehouse tablets, causing UI lag and dropping canvas draw frames.
2. Google MediaPipe Hands
MediaPipe Hands uses an ML pipeline consisting of a palm detector and a hand landmark model mapping 21 3D coordinates. MediaPipe runs inside WebAssembly (Wasm) with hardware acceleration. It consumed less than 15% CPU overhead on the same hardware, maintained stable hand tracking under heavy backlighting, and provided depth (Z-axis) coordinates, which allowed us to track the "press" (hover) state of the pen.
Project Metrics & Impact
- Hardware Support: 10,000+ low-spec Celeron-based warehouse terminals.
- CPU Usage: Reduced from 85% (Custom CNN) to under 15% (MediaPipe Wasm).
- Drawing Latency: Less than 8ms loop time, maintaining 30 FPS canvas rendering.
- Signatures Captured: 3,000+ secure, contactless signatures collected.
- Data Security: 100% private. All coordinates processed locally; zero frames uploaded.
Implementing AirSign: Jitter, Smoothing, and PDF Signing
To translate the finger-tip tracking into a clean, legal signature, our engineering team resolved two key problems:
1. Smoothing Hand Jitter
Physical hand tracking contains micro-jitters due to fatigue. If plotted directly, the signature looks jagged. We implemented a double-exponential smoothing algorithm (Kalman Filter) on the incoming coordinate stream. The filter computes the velocity of hand movement and dynamically adjusts the smoothing factor: stronger smoothing during slow strokes, and rapid response during quick sweeps.
2. Client-Side Cryptographic Signatures
Rather than sending the signature image to a server to burn onto a document, we compile the certificate entirely in the browser. Using the Forge-PDF library, the client stamps the SVG drawing vectors, attaches local device telemetry (terminal ID, timestamp), generates a secure hash, and compiles the final certified PDF. This guarantees that Velo Logistics has 100% data isolation and complies with enterprise compliance protocols.
Conclusion
By prioritizing client-side calculations and utilizing MediaPipe's Wasm model, Ananta Labs delivered a robust contactless signature engine. Velo Logistics saved thousands in terminal hardware maintenance fees and established a modern touchless workflow.