TechPrufer
Get started
Grab
Grab logoGrab · Feature trail

Facial Recognition System

2 posts · Jun 2022 – Nov 2025

What this trail covers

Grab has developed and implemented a comprehensive facial recognition system for various applications, including identity verification for consumers and partners, and electronic Know Your Customer (e-KYC) processes for financial services. The system comprises image preprocessing (detection and alignment), face anti-spoofing mechanisms to detect screen replay and printed photo attacks, and feature extraction to generate distinguishing embeddings. Downstream applications include face verification. This post details the development of a custom, lightweight Vision LLM to improve document processing, specifically for eKYC, by overcoming limitations of traditional OCR and proprietary LLMs. The custom model, built from scratch, achieves comparable accuracy to larger models with significantly reduced latency and improved SEA language support. It involves a multi-stage training process including projector alignment, vision tower enhancement, language-specific visual training, and task-centric fine-tuning.

All Grab trails
  1. How we built a custom vision LLM to improve document processing at GrabThis post details the development of a custom, lightweight Vision LLM for document processing, specifically for eKYC. It describes the limitations of existing OCR and proprietary LLMs for SEA languages and the decision to build a specialized model. The process involved selecting Qwen2-VL 2B as a base, generating synthetic OCR datasets, and developing an auto-labeling framework (Documint). Experiments with LoRA fine-tuning showed promise but limitations with non-Latin scripts. Full fine-tuning of Qwen2VL-2B significantly improved accuracy. Ultimately, a custom ~1B parameter Vision LLM was built from scratch by combining components from Qwen2-VL 2B and Qwen2.5 0.5B, trained in four stages: projector alignment, vision tower enhancement, language-specific visual training, and task-centric fine-tuning. The custom model achieved comparable accuracy to the 2B model with significantly reduced latency and improved SEA language support.
  2. How facial recognition technology keeps you safeThis post details the technical implementation of Grab's facial recognition system. It covers the application of facial recognition for identity verification and e-KYC, and elaborates on the technical pipeline including face detection, alignment, anti-spoofing (with techniques for synthetic data generation and patch-based training), and feature extraction. It also discusses face verification, including the challenges of selfie-to-ID and selfie-to-masked selfie scenarios, and the use of semi-Siamese training and masked selfie data for model improvement.