Forward Pass

import * as tf from '@tensorflow/tfjs';

// Step 1: Tokenize
function tokenize(text: string): string[] {
return text.toLowerCase().replace(/[.,!?;:]/g, ' ').split(/s+/).filter(Boolean);
}

// Step 2: Encode to bag-of-words vector
function toBagOfWords(tokens: string[], vocab: Map<string, number>, vocabSize: number): number[] {
const bow = new Array(vocabSize).fill(0);
for (const token of tokens) {
  const idx = vocab.get(token);
  if (idx !== undefined) bow[idx] += 1;
}
return bow;
}

// Step 3: Classify
function forwardPass(text: string, vocab: Map<string, number>,
                   weights: tf.Tensor, bias: tf.Tensor): number {
const tokens = tokenize(text);
const bow = toBagOfWords(tokens, vocab, vocab.size);
const input = tf.tensor(bow);
const score = input.mul(weights).sum().add(bias);
return tf.sigmoid(score).dataSync()[0];
}

// Build vocabulary
const vocab = new Map<string, number>([
['patient', 0], ['allergy', 1], ['penicillin', 2],
['medication', 3], ['reports', 4], ['pain', 5],
['discharge', 6], ['follow', 7],
]);

// Trained weights (in practice, these come from training)
const weights = tf.tensor([0.1, 0.9, 0.8, 0.3, 0.0, -0.1, -0.5, -0.3]);
const bias = tf.scalar(-0.2);

// Run the pipeline
const note = 'Patient has documented penicillin allergy';
const probability = forwardPass(note, vocab, weights, bias);
console.log('Allergy probability:', probability.toFixed(3)); // ~0.83

// Frontend pipeline (you know this)
const result = await fetch('/api/notes/42')
.then(res => res.json())           // parse
.then(data => validate(data))       // validate
.then(valid => renderNote(valid));   // render

// ML pipeline (same pattern!)
const prediction = pipe(
tokenize,                            // parse text
tokens => toBagOfWords(tokens, vocab, vocab.size),  // vectorize
bow => tf.tensor(bow).mul(weights).sum().add(bias), // compute
score => tf.sigmoid(score).dataSync()[0],           // activate
)('Patient has documented penicillin allergy');