Back to Home
Engineering Works
Low-level neural algorithms, secure LAN infrastructure, and real-time state engines.
3 Repositories
FeedforwardNN
A feedforward neural network implemented from scratch in C++ to demonstrate the math behind forward and backward passes.
C++PythonMake
Technical Implementation
- Architecture: 2 input → 1 hidden layer (configurable) → 1 output
- Activation: Sigmoid (hidden + output)
- No bias terms
- Weight initialization: Random
- Training: Online / stochastic gradient descent
- Loss: Absolute error (averaged, logged every 100 epochs)
- Training length: 10,000 epochs
- Learning rate: 0.1
- Logging: CSV output + Python loss plot
- Purpose: clarity over production features
snippet.cppVerified
// Neural Network Forward Pass in C++
void ForwardPass(const Matrix& input) {
hidden_layer = Sigmoid(input * w1 + b1);
output_layer = Sigmoid(hidden_layer * w2 + b2);
loss = ComputeLoss(target, output_layer);
}Tone: mathematical, instructional, honestView Repository on GitHub
Next.jsTypeScriptWebCrypto API
Engineering Emphasis
- Client-side cryptography
- Per-recipient key packing
- Server never sees plaintext or raw symmetric keys
- In-memory server + heartbeat polling
- LAN discovery without heavy infra
- LocalStorage persistence
- Explicit client-side key management
snippet.tsVerified
// End-to-End Client Encryption
const key = await crypto.subtle.generateKey(
{ name: "AES-GCM", length: 256 },
true, ["encrypt", "decrypt"]
);
const encrypted = await crypto.subtle.encrypt(
{ name: "AES-GCM", iv }, key, payload
);Tone: security-aware, system-design focusedView Repository on GitHub
Socket.IONode.jsReact
snippet.tsVerified
// Authoritative Server State Sync
io.on("connection", (socket) => {
socket.on("submit_word", (payload) => {
const valid = validateWordState(gameRoom, payload);
if (valid) io.to(gameRoom.id).emit("state_update", gameRoom.state);
});
});Tone: event-driven, real-time systems thinkingView Repository on GitHub