Fest Ticketing AI

Overview

Fest Ticketing AI was built as a combined project for 3 university courses (Digital Image Processing, Machine Learning, and Mobile Programming). The system uses face recognition for event ticketing — attendees register their face, and at the venue, a camera captures their image and matches it against the database. I built the AI backend pipeline using FastAPI, MTCNN for face detection, Facenet for embedding generation, and Cosine Similarity for matching.

Problem

Traditional ticketing systems can be bypassed with stolen or shared tickets
The coursework required integrating AI (face recognition) with a mobile application
Training a face recognition model from scratch was not feasible within the deadline and hardware constraints

Goal

Build a backend API that handles face detection, embedding, and matching
Integrate pre-trained models (MTCNN + Facenet) instead of training from scratch
Create a working PoC that demonstrates end-to-end face recognition for ticketing

Role

I built the AI backend pipeline independently, working alongside a mobile team (Flutter) and a frontend team.

Team: Achmad Raihan Fahrezi Effendy (AI Backend Engineer) + Mobile Team + Frontend Team

Architecture

flowchart TD A["Flutter Mobile App"] --> B["FastAPI Backend"] B --> C["Face Detection"] B --> D["Face Embedding"] B --> E["Face Matching"] C --> F["MTCNN"] D --> G["Facenet Pre-trained"] E --> H["Cosine Similarity"] B --> I[("PostgreSQL / MySQL")] J["OpenCV"] --> C K["MediaPipe"] --> C

Layer	Technology	Purpose
Backend	FastAPI (Python)	Async API server
Face Detection	MTCNN	Detect faces in images
Face Embedding	Facenet (pre-trained)	Generate face vectors
Matching	Cosine Similarity	Compare face embeddings
Database	PostgreSQL / MySQL	Store face embeddings and user data
Image Processing	OpenCV, MediaPipe	Image preprocessing
Container	Docker	Reproducible environment

Key Features

Face Detection — MTCNN detects faces in uploaded images with bounding box coordinates
Face Embedding — Facenet generates 128-dimensional face vectors for comparison
Face Matching — Cosine Similarity compares embeddings to identify registered users
Liveness Detection — Mobile app captures multiple angles (front, left, right, smile) as a simple liveness check
Async Processing — FastAPI handles concurrent face recognition requests efficiently

Technical Decisions

FastAPI was chosen for its async support and Python ecosystem for AI/ML libraries
Pre-trained models (MTCNN + Facenet) were used instead of training from scratch due to time and hardware constraints
Cosine Similarity was selected for face matching because of its simplicity and effectiveness
PostgreSQL was used for storing face embeddings with vector operations

Challenges

Training a model from scratch failed due to time and hardware limitations — pivoting to pre-trained models was a critical decision
The face recognition pipeline needed to handle varying lighting conditions and angles
Integrating the mobile app (Flutter) with the Python backend required careful API design

Outcome

The system successfully demonstrated end-to-end face recognition for event ticketing. Users could register their face via the mobile app, and the backend could identify them at the venue. The working PoC proved the concept was viable, though it was not production-ready.

Final Thoughts

Fest Ticketing AI taught me that working PoC is better than perfect theory. The pivot from training custom models to using pre-trained models was not a compromise — it was a practical decision that delivered results within constraints. The experience also showed me how different技术栈 (Python AI backend + Flutter mobile) can integrate through well-designed APIs.