Lab 12: AI in the Real World — Healthcare, Finance, Security

Objective

See how AI is deployed across industries today. By the end you will understand:

• Concrete AI applications in healthcare, finance, and cybersecurity

• The gap between research benchmarks and production realities

• Key failure modes and why responsible deployment matters

• Career paths and opportunity areas in applied AI

---

Healthcare: AI as a Diagnostic Tool

Medical Imaging

AI has reached or exceeded radiologist-level accuracy on several imaging tasks:

Task	Model	Performance	vs. Human
Diabetic retinopathy screening	Google DeepMind	94.5% AUC	Comparable
Lung cancer detection (CT)	Google Healthcare AI	Surpasses 6/6 radiologists	Better
Breast cancer screening (mammography)	DeepMind	11.5% fewer false positives	Better
Skin cancer classification	Stanford CNN	Dermatologist-level on 21 conditions	Comparable

# Conceptual: medical image classifier using transfer learning
import torchvision.models as models
import torch.nn as nn

class ChestXRayClassifier(nn.Module):
    """Classify chest X-rays into 14 pathologies (CheXNet architecture)"""

    def __init__(self, num_classes=14):
        super().__init__()
        # DenseNet-121 pre-trained on ImageNet
        self.densenet = models.densenet121(pretrained=True)

        # Replace final layer for multi-label classification
        num_features = self.densenet.classifier.in_features
        self.densenet.classifier = nn.Sequential(
            nn.Linear(num_features, num_classes),
            nn.Sigmoid()    # multi-label: each class independent 0-1 probability
        )

    def forward(self, x):
        return self.densenet(x)

# Output: [0.92, 0.03, 0.87, ...]
# Labels: [Atelectasis, Cardiomegaly, Effusion, ...]
# Radiologist reviews AI flags — AI is the triage, human is the decision

Drug Discovery

DeepMind's AlphaFold 2 (2020) solved the 50-year protein folding problem — predicting a protein's 3D structure from its amino acid sequence. This matters because protein structure determines function and therefore drug targets.

• 200 million protein structures predicted (vs. ~150,000 known experimentally)

• Used in research against malaria, Parkinson's, antibiotic resistance

• Nobel Prize in Chemistry awarded to Demis Hassabis (2024)

Clinical Language Understanding

# Clinical NLP: extract structured data from clinical notes
prompt = """
Extract the following from this clinical note and return JSON:
- diagnoses (list)
- medications (list with dosage)
- lab values (dict)
- follow_up_required (boolean)

Note: Patient presents with T2DM poorly controlled (HbA1c 9.2%). 
Current medications: Metformin 1000mg BD, Empagliflozin 10mg OD. 
Starting Ozempic 0.25mg weekly. Follow up in 3 months for repeat HbA1c.
"""
# AI extracts structured data from unstructured clinical text
# Feeds into EHR systems, billing, population health dashboards

---

Finance: AI in Markets, Risk, and Fraud

Algorithmic Trading

Hedge funds like Renaissance Technologies, Two Sigma, and D.E. Shaw have used ML for decades. Modern approaches:

import numpy as np

# Feature engineering for stock price prediction
def build_features(price_history: np.ndarray, volume: np.ndarray) -> np.ndarray:
    """Technical indicators as ML features"""
    features = {}

    # Moving averages
    features['sma_5']  = np.convolve(price_history, np.ones(5)/5,  mode='valid')[-1]
    features['sma_20'] = np.convolve(price_history, np.ones(20)/20, mode='valid')[-1]

    # Momentum
    features['roc_5']  = (price_history[-1] - price_history[-6])  / price_history[-6]
    features['roc_20'] = (price_history[-1] - price_history[-21]) / price_history[-21]

    # Volatility
    returns = np.diff(np.log(price_history[-21:]))
    features['volatility_20'] = returns.std() * np.sqrt(252)

    # Volume ratio
    features['volume_ratio'] = volume[-1] / volume[-20:].mean()

    return np.array(list(features.values()))

# These features feed into gradient boosting, LSTM, or Transformer models
# Predicting: direction (up/down), magnitude, or optimal position size

Credit Scoring

Traditional credit scoring (FICO) uses 5 variables. ML uses thousands:

• Traditional features: payment history, credit utilisation, length of history

• Alternative features: bank transaction patterns, employment history, education, social connections

Regulatory concern: Alternative features can be proxies for protected characteristics (race, gender). Regulators in UK (FCA), EU (ECB), and US (CFPB) require explainability for credit decisions.

# SHAP: explain why a credit application was declined
import shap

explainer = shap.TreeExplainer(credit_model)
shap_values = explainer.shap_values(applicant_features)

# Output: which features pushed the score up or down
# "Your application was declined because:
#  - Short credit history: -45 points
#  - High credit utilisation (82%): -38 points
#  - Recent missed payment: -52 points"

Fraud Detection

Payment fraud detection must operate at millisecond speed with ~0.1% false positive rate (or customers complain):

# Real-time fraud scoring (conceptual)
def score_transaction(transaction: dict) -> float:
    """Return fraud probability 0-1"""
    features = extract_features(transaction)
    # Features include:
    # - Transaction amount vs. historical average
    # - Geographic distance from last transaction
    # - Time since last transaction  
    # - Merchant category risk score
    # - Device fingerprint match
    # - Velocity: transactions in last hour/day
    
    score = fraud_model.predict_proba([features])[0][1]
    
    if score > 0.95:   block_transaction()
    elif score > 0.7:  request_authentication()
    else:              approve_transaction()
    
    return score

Visa processes 65,000+ transactions per second — each scored by ML in <100ms.

---

Cybersecurity: AI Attacks and Defences

Threat Detection

# Network anomaly detection using Isolation Forest
from sklearn.ensemble import IsolationForest
import pandas as pd

# Features: network packet statistics
features = pd.DataFrame({
    'bytes_sent':    [1200, 1100, 5_000_000, 1300],  # 3rd = suspicious
    'packet_count':  [12,   11,   45000,    13],
    'unique_dests':  [3,    2,    847,      4],
    'duration_sec':  [2.1,  1.8,  300.0,   2.3],
})

# Isolation Forest: anomalies are easier to isolate (shorter path in tree)
model = IsolationForest(contamination=0.01, random_state=42)
model.fit(features)
predictions = model.predict(features)
# -1 = anomaly, 1 = normal
# Row 3 (index 2): flagged as anomaly — potential data exfiltration

AI-Generated Malware (The Threat)

LLMs have lowered the barrier to creating sophisticated malware:

• 2023: Researchers demonstrated GPT-4 writing functional phishing emails, polymorphic code, and social engineering scripts

• WormGPT/FraudGPT: Uncensored LLMs sold on darknet forums specifically for cybercrime

• AI-generated spear phishing: Personalised to target's LinkedIn, reducing detection by humans

Deepfakes and Social Engineering

Deepfake audio attack (2019, UK): Criminals used AI voice cloning to impersonate a German executive's voice on a phone call, directing a UK employee to transfer €220,000. The voice was indistinguishable.

Scale of the problem (2024):

• 25 million deepfake video/audio pieces circulating online

• 3,000% increase in deepfake fraud attempts (2023→2024)

• AI-generated fake IDs passing document verification systems

AI for Defence

# LLM-assisted code security review
code_review_prompt = """
Review this code for security vulnerabilities. 
For each vulnerability:
1. Name the vulnerability type (e.g., SQL injection, XSS)
2. Identify the specific line(s)
3. Explain the attack vector
4. Provide the remediated code

Code to review:
{code}
"""

# Deployed by:
# - GitHub Advanced Security (CodeQL + LLM explanations)
# - Snyk Code (AI-powered SAST)
# - Amazon CodeGuru (automated code review)

---

Legal, Education, and Creative Industries

Legal

• Contract review: Kira Systems, Harvey AI — review 1,000-page contracts in minutes

• Legal research: Lexis+ AI, Westlaw AI — cite cases, find precedents

• E-discovery: AI reviews millions of documents for relevant evidence

Education

• Khan Academy Khanmigo — personalised Socratic tutor (never just gives answers)

• Duolingo Max — AI conversation partner for language learning

• GitHub Copilot — used by 1.3M developers; 55% of code in some repos is AI-generated

Creative Industries

Tool	Use Case	Controversy
GitHub Copilot	Code generation	Trained on GPL code without licence compliance
Midjourney	Image generation	Artists sue over training data
Suno/Udio	Music generation	Record labels sue for copyright
ElevenLabs	Voice cloning	Voice actors sue; used for fraud

---

The Production Reality Gap

Research benchmarks vs. production systems:

Research	Production Reality
95% accuracy	95% accuracy on test set, 73% on real users (distribution shift)
English only	Degrades significantly for non-English speakers
Clean images	Struggles with phone photos: blur, glare, unusual angles
Single task	Users ask off-topic questions constantly
Offline batch	Must respond in <500ms at 10,000 requests/second

The last mile from "works in research" to "works reliably in production" is where most AI projects fail.

---

Career Paths in Applied AI

Role	Skills	Typical Entry
ML Engineer	Python, PyTorch/TensorFlow, MLOps, cloud	CS degree or bootcamp + portfolio
Data Scientist	Python, statistics, SQL, business insight	Stats/CS degree
AI Product Manager	Domain expertise + AI literacy	PM experience + AI upskilling
AI Safety Researcher	Deep learning theory, philosophy, formal methods	PhD level
Prompt Engineer	LLM expertise, domain knowledge	Emerging role; no formal path
MLOps Engineer	Docker, Kubernetes, monitoring, CI/CD	DevOps → ML

---

Further Reading

• AlphaFold and the Protein Folding Revolution

• AI in Finance — BIS Working Papers

• AI and Cybersecurity — ENISA Report

• The State of AI Report 2024