the one that started it all
September 19, 2004 · Windows XP · original screenshot
Face Recognition
using Neural Networks
A real-time biometric security system. Captures a face from a digital camera via TWAIN, extracts features using PCA (eigenfaces), and matches against a database using a multi-layer neural network. The bar chart shows confidence scores per person. The network topology is visualized live. Written in raw C++ with Win32 API — no frameworks, no libraries, no GPU. Three people, one semester, one compiler.
↓ face_rec_source.zip (4.1 MB — full source, training data & saved network weights)Twenty-one years later, the same person runs teams of AI agents that build software autonomously. The arc is longer than you think.
live: the algorithm itself, faithfully
Running the actual 2003 pipeline on the actual training set.
The Edgeeye.dat / Pixeye.dat networks (4 MB of weights from 2003) are loaded by a Python port of the original C++ pipeline, run against each of the 8 training photos as input, and the resulting numbers are played back here. The demo cycles through all 8 test inputs — watch what actually happens vs. what should.
8 tests, 8 correct — the algorithm works. For some photos the 80×50 “eye box” lands on cheeks (look at Gowtham and Rajesh); that's the real algorithm in action — it doesn't insist on the anatomical eye, it finds whatever 80×50 patch the pre-trained edge+pixel nets recognise. Eye-detection here matches the original Feature.dat values to the pixel for several photos.
