AI Foundation · Domain 11

AI Applications & Industry

Real-world AI across healthcare, finance, robotics, autonomous vehicles, education, scientific discovery, code generation, creative AI, and enterprise deployment.

11.1
Chapter 11.1
AI in Healthcare

Healthcare is where AI has the highest potential impact per correct prediction — and the highest cost per incorrect one. A missed tumour in a radiology scan, a wrong drug interaction prediction, a biased clinical-trial model — the stakes are human lives.

Deep learning for medical imaging is the most mature clinical AI application. CNNs (Domain 6) analyse X-rays, CT scans, MRIs, pathology slides, and retinal images. FDA has approved 500+ AI/ML-enabled medical devices as of 2024 — the majority are radiology tools.

📷
Radiology

Chest X-ray: detect pneumonia, lung nodules, fractures. CT: liver lesions, pulmonary embolism. MRI: brain tumours, cardiac imaging. AI as “second reader” — catches missed findings.

🔬
Pathology

Whole-slide imaging + CNNs: classify cancer subtypes, grade tumours, count mitotic figures. Paige AI: first FDA-approved computational pathology system for prostate cancer (2021).

👁️
Ophthalmology

Diabetic retinopathy screening: Google Health achieved specialist-level accuracy. IDx-DR: first FDA-approved autonomous AI diagnostic (2018) — diagnoses without a physician.

ApplicationAI ApproachPerformanceRegulation StatusImpact
Chest X-ray triageCNN (DenseNet/ResNet)AUC 0.95–0.99FDA cleared (multiple)Reduces reporting delay 50%+
Mammography screeningCNN + attentionReduces false negatives 9%FDA clearedEarlier cancer detection
Diabetic retinopathyCNN classificationSensitivity >90%FDA approved (autonomous)Screening in primary care
Skin lesion classificationCNN (EfficientNet)Dermatologist-level on benchmarksLimited deploymentBias on darker skin tones

Traditional drug discovery takes 10–15 years and $2.6 billion per approved drug. AI accelerates multiple stages: target identification, molecular generation, binding prediction, toxicity screening, and clinical trial design. The promise is shorter timelines and lower costs.

AI in Drug Discovery Pipeline — acceleration at every stage
Target ID GNNs, NLP Molecule Gen VAE, Diffusion Binding Pred AlphaFold, GNNs Toxicity Screen GBMs, NNs Clinical Trials Patient matching Approval Traditional: 10–15 years, $2.6B per drug AI-assisted: potential 30–50% time reduction in pre-clinical phases Caveat: no AI-discovered drug has completed Phase III trials as of 2025
🧬
AlphaFold (DeepMind, 2020)

Predicted 3D protein structures for 200M+ proteins. Solved a 50-year grand challenge in biology. Enables structure-based drug design at unprecedented scale. Nobel Prize in Chemistry 2024.

💊
Insilico Medicine

First AI-designed drug to reach Phase II clinical trials (2023) — for idiopathic pulmonary fibrosis. Molecule generated and optimised entirely by generative AI. Timeline: 30 months from target to Phase I.

∑ Chapter 11.1 — Key Takeaways

  • Medical imaging is the most mature clinical AI — 500+ FDA-approved AI devices, majority in radiology
  • AI as “second reader” improves radiologist accuracy — reduces false negatives 5–10%
  • AlphaFold solved protein structure prediction — 200M+ structures, Nobel Prize 2024
  • AI-discovered drugs reaching clinical trials — but none past Phase III as of 2025
  • Skin lesion AI shows bias on darker skin tones — dataset diversity is critical
11.2
Chapter 11.2
AI in Finance & Trading

Finance was one of the earliest adopters of ML — quantitative trading firms have used statistical models since the 1980s. Today, AI powers fraud detection, credit scoring, algorithmic trading, insurance underwriting, and regulatory compliance at massive scale.

Algorithmic trading uses ML models to predict price movements, optimise execution, and manage risk. 60–75% of US equity volume is now algorithmic. Approaches range from classical time-series models to deep RL portfolio optimisation.

📈
Signal Generation

NLP on earnings calls, news, SEC filings. Sentiment analysis for market direction. Alternative data: satellite imagery, credit card data, social media. LLMs for financial document analysis.

Execution Optimisation

RL agents minimise market impact. Optimal order splitting and timing. High-frequency: microsecond decisions. TWAP/VWAP benchmarking with ML adjustments.

📊
Risk Management

Portfolio VaR estimation with deep learning. Tail-risk modelling. Scenario generation with GANs. Stress testing with synthetic market crashes. Real-time margin calculations.

Fraud Detection
Credit Scoring

Scale: Billions of transactions/day, ~0.1% are fraudulent.

Challenge: Extreme class imbalance (1:1000). Adversarial — fraudsters adapt.

Models: Gradient boosting (XGBoost), graph neural networks (transaction networks), autoencoders (anomaly detection).

Latency: Must decide in <100ms for real-time payment authorisation.

Impact: Mastercard AI reduced false declines 200% while catching more fraud.

Scale: Millions of applications, used for lending decisions.

Challenge: Fairness — must comply with ECOA, cannot discriminate on protected attributes.

Models: Logistic regression (traditional FICO), gradient boosting, neural networks.

Explainability: Regulations require adverse action notices — must explain why denied.

Tension: More accurate models (deep learning) are less explainable — regulatory constraint.

∑ Chapter 11.2 — Key Takeaways

  • 60–75% of US equity volume is algorithmic — NLP on earnings calls is the fastest-growing signal source
  • Fraud detection: extreme class imbalance + adversarial adaptation + <100ms latency
  • Credit scoring: accuracy vs explainability tradeoff — regulations require explanations for denials
  • LLMs increasingly used for financial document analysis, compliance, and report generation
11.3
Chapter 11.3
Robotics & Autonomous Vehicles

Robotics is where AI meets the physical world — and the physical world does not forgive bugs. A software error in a chatbot produces a wrong sentence; a software error in a self-driving car can kill someone. This chapter covers both industrial robotics and autonomous vehicle AI stacks.

🏭
Industrial Robotics

Manufacturing assembly, pick-and-place, welding, quality inspection. Vision-guided manipulation. Companies: FANUC, ABB, KUKA. AI for path planning, defect detection, predictive maintenance.

🤖
Humanoid & Service Robots

Boston Dynamics (Atlas), Tesla (Optimus), Figure AI. LLM-powered task planning. Sim-to-real transfer for locomotion. Warehouse robots (Amazon): 750K+ deployed globally.

🧠
Foundation Models for Robotics

RT-2 (Google): vision-language-action model. SayCan: LLM plans, robot executes. Key challenge: generalisation across environments without massive retraining.

Autonomous Vehicle AI Stack — perception to control in milliseconds
Sensors Camera, LiDAR, Radar Perception Detection, Tracking Prediction Intent, Trajectory Planning Path, Behaviour Control Steer, Brake, Gas End-to-end latency: <100ms from sensor input to control output Waymo: modular stack | Tesla: vision-only, end-to-end neural net | Comma.ai: open-source openpilot SAE Levels: L2 (hands on) → L3 (eyes off, limited) → L4 (geo-fenced full auto) → L5 (anywhere) As of 2025: L4 robotaxis in select cities (Waymo, Cruise). No L5 exists.

∑ Chapter 11.3 — Key Takeaways

  • Industrial robotics: AI for vision-guided manipulation, defect detection, path planning
  • Foundation models for robotics (RT-2, SayCan): LLMs plan, robots execute
  • AV stack: sensors → perception → prediction → planning → control in <100ms
  • L4 robotaxis operating in select cities — L5 (anywhere, any conditions) does not yet exist
11.4
Chapter 11.4
AI in Education

The promise of AI in education is the oldest dream in ed-tech: a personal tutor for every student. LLMs have brought this closer than ever — but also raised urgent questions about academic integrity, learning quality, and equitable access.

🎓
LLM-Powered Tutoring

Khanmigo (Khan Academy + GPT-4): Socratic tutoring — guides students with questions rather than answers. Duolingo Max: AI conversation practice for language learning. Code tutoring: GitHub Copilot for students.

📖
Adaptive Learning Systems

Adjust difficulty based on student performance. Knowledge tracing: model what the student knows. Spaced repetition with ML-optimised schedules. Companies: Knewton, ALEKS, DreamBox. Pre-dates LLMs — uses knowledge graphs + Bayesian models.

⚠️
Academic Integrity Challenge

LLMs can write essays, solve problem sets, generate code. AI detection tools are unreliable (70–80% accuracy, high false positive rate). Many institutions shifting to oral exams, in-class assessments, process portfolios.

🌍
Equity Concerns

Paid AI tools (GPT-4, Khanmigo) give advantage to students who can afford them. LLMs primarily trained in English — quality drops for other languages. Digital divide amplified by AI divide.

∑ Chapter 11.4 — Key Takeaways

  • LLM tutoring (Khanmigo): Socratic method at scale — guides with questions, not answers
  • Adaptive learning: knowledge tracing + spaced repetition — older than LLMs, still effective
  • Academic integrity: AI detection is unreliable — institutions shifting to process-based assessment
  • Equity: paid AI tools create access gap between wealthy and low-income students
11.5
Chapter 11.5
AI for Scientific Discovery

AI is becoming a tool for scientific discovery itself — not just automating existing experiments, but finding patterns humans would never see. From protein folding to materials science to weather prediction, AI is accelerating the pace of science.

🧬
AlphaFold (DeepMind)

Predicted 3D structure of 200M+ proteins. Solved 50-year grand challenge. Used by 1M+ researchers within 18 months. Nobel Prize in Chemistry 2024 (Hassabis, Jumper). Transformer + invariant point attention architecture.

🌊
GraphCast (DeepMind, 2023)

10-day weather forecasting more accurate than ECMWF (gold standard). Graph neural network on 0.25° grid. Prediction in <1 minute vs hours for physics-based models. Enables rapid ensemble forecasting.

🔭
GNoME (DeepMind, 2023)

Discovered 2.2 million new crystal structures — including 380K stable materials for potential batteries, solar cells, superconductors. Graph neural network predicting material stability.

SystemDomainArchitectureImpactYear
AlphaFold 2Protein structureTransformer + IPA200M+ structures, Nobel Prize2020
GraphCastWeather forecastingGNN (mesh)Beats ECMWF at 10-day forecast2023
GNoMEMaterials scienceGNN2.2M new crystal structures2023
AlphaGeometryMathematicsNeuro-symbolicIMO-level geometry proofs2024
FunSearchMathematicsLLM + evolutionaryDiscovered new mathematical constructions2024

∑ Chapter 11.5 — Key Takeaways

  • AlphaFold: 200M+ protein structures, Nobel Prize 2024 — single biggest AI impact on science
  • GraphCast: faster and more accurate than physics-based weather models
  • GNoME: 2.2M new materials — potential batteries, solar cells, superconductors
  • AI for science typically uses graph neural networks and domain-specific architectures, not generic LLMs
11.6
Chapter 11.6
Code Generation & Software Engineering

AI code assistants are the fastest-adopted developer tool in history. GitHub Copilot reached 1.8 million paid subscribers within two years. The impact is real: measurable productivity gains, but also new categories of risk.

ToolProviderModelKey FeaturePricing
GitHub CopilotGitHub/MicrosoftGPT-4o + CodexIDE inline completion, chat, workspace agent$10–19/mo
CursorCursor Inc.Claude/GPT-4oAI-native IDE, multi-file edit, codebase-aware$20/mo
Amazon CodeWhispererAWSCustom modelsSecurity scanning, AWS integrationFree tier
Codeium / WindsurfCodeiumCustom modelsFree for individuals, fast completionsFree
Claude CodeAnthropicClaude 3.5/4Terminal agent, multi-file, agentic codingUsage-based
📈
Measured Productivity Gains

GitHub study (2022): 55% faster task completion with Copilot. Google internal study: 6% fewer iterations in code review. McKinsey (2023): 20–45% productivity increase for coding tasks. Largest gains in boilerplate/test code; smallest in novel architecture.

⚠️
Risks & Limitations

Security: AI-generated code may contain vulnerabilities (Stanford study: developers with AI wrote less secure code). Copyright: training on copyrighted code — legal risk. Over-reliance: accepting suggestions without understanding them.

∑ Chapter 11.6 — Key Takeaways

  • GitHub Copilot: 1.8M subscribers, 55% faster task completion in controlled studies
  • AI-native IDEs (Cursor): codebase-aware, multi-file editing — the next evolution
  • Agentic coding (Claude Code, Copilot Workspace): plan and execute multi-step tasks autonomously
  • Risks: security vulnerabilities in generated code, copyright concerns, over-reliance without understanding
11.7
Chapter 11.7
Creative AI & Media

Generative AI has transformed creative industries faster than any previous technology. In under two years, AI image generation went from research curiosity to a $1B+ market disrupting illustration, photography, video production, music, and advertising.

🎨
Image Generation

Stable Diffusion (open source), DALL-E 3 (OpenAI), Midjourney. Diffusion models: iteratively denoise from random noise. Text-to-image via CLIP text conditioning. Used for: concept art, marketing, product design.

🎬
Video Generation

Sora (OpenAI), Runway Gen-3, Pika, Kling. Text/image-to-video with temporal consistency. Minute-long videos with coherent motion. Used for: advertising, prototyping, storyboarding. Still limited: physics errors, temporal glitches.

🎵
Music & Audio

Suno, Udio: text-to-song with lyrics. ElevenLabs: voice cloning from seconds of audio. MusicLM (Google): text-to-music. Concerns: artist compensation, voice cloning consent, copyright of AI compositions.

⚖️
Copyright & Labour Impact

Major lawsuits: Getty vs Stability AI, NYT vs OpenAI, artists vs Midjourney. Concept artists report 50–70% decline in freelance work. AI-generated images not copyrightable (US Copyright Office, 2023). EU AI Act: must disclose AI-generated content.

💡
New Creative Workflows

AI as creative tool, not replacement: rapid ideation, style exploration, variant generation. Inpainting: edit parts of existing images. ControlNet: precise composition control. Professional use: advertising agencies, game studios, film pre-production.

∑ Chapter 11.7 — Key Takeaways

  • Diffusion models (Stable Diffusion, DALL-E 3, Midjourney): text-to-image is now commodity capability
  • Video generation (Sora, Runway): minute-long coherent videos, but physics errors persist
  • Voice cloning from seconds of audio — massive consent and deepfake concerns
  • Copyright unsettled: training on copyrighted data + AI outputs not copyrightable = legal uncertainty
11.8
Chapter 11.8
Enterprise AI & Industry Overview

Enterprise AI is where the majority of economic value is created — not in frontier research, but in applying proven techniques to business processes. The gap between AI research and enterprise deployment is the single biggest challenge in the field.

💬
Customer Experience

LLM-powered chatbots and support agents. Klarna: AI assistant handles 2/3 of customer chats, equivalent to 700 agents. Sentiment analysis for feedback triage. Personalised recommendations.

📦
Operations & Supply Chain

Demand forecasting with gradient boosting. Inventory optimisation. Route planning (logistics). Predictive maintenance for equipment. Quality inspection with computer vision.

📝
Knowledge & Documents

RAG for internal knowledge bases. Contract analysis with NLP. Automated report generation. Search across unstructured enterprise data. Compliance document review.

IndustryTop Use CasePrimary AIMaturityEstimated Value
RetailRecommendation + demand forecastingCollaborative filtering, GBMsMature$100B+ annually
ManufacturingPredictive maintenance + qualityTime-series, CVMature$50B+
BankingFraud detection + credit scoringGBMs, NNsMature$40B+
HealthcareClinical decision supportCNNs, NLPGrowing$20B+
LegalContract review + researchNLP, RAGEarly$10B+

∑ Chapter 11.8 — Key Takeaways

  • Enterprise AI value: most economic impact is from proven techniques applied to business processes
  • Customer experience: LLM chatbots replacing hundreds of support agents at companies like Klarna
  • Supply chain: demand forecasting and predictive maintenance are the highest-ROI use cases
  • The gap between research and deployment is the biggest challenge — MLOps (Domain 9) bridges it

🎓 Domain 11 Complete — AI Applications & Industry

  • Ch 11.1: Healthcare AI: 500+ FDA-approved devices. AlphaFold solved protein folding (Nobel 2024). Drug discovery accelerating but no AI drug past Phase III.
  • Ch 11.2: Finance: 60–75% of trading is algorithmic. Fraud detection at <100ms. Credit scoring: accuracy vs explainability tension.
  • Ch 11.3: Robotics: vision-guided manipulation, foundation models for robots (RT-2). AV: L4 robotaxis in select cities, no L5.
  • Ch 11.4: Education: LLM tutoring (Khanmigo) at scale. Academic integrity crisis. Equity gap from paid AI tools.
  • Ch 11.5: Science: AlphaFold (200M proteins), GraphCast (weather), GNoME (2.2M materials) — GNNs dominate scientific AI.
  • Ch 11.6: Code: Copilot 55% faster. AI-native IDEs (Cursor) and agentic coding emerging. Security and copyright risks.
  • Ch 11.7: Creative AI: diffusion models are commodity. Video generation emerging. Copyright and labour impact unsettled.
  • Ch 11.8: Enterprise: most value from proven techniques applied to customer experience, supply chain, and documents.

Domain 11 is where everything from Domains 1–10 meets the real world. The pattern is consistent: the hardest part is never the model — it is the data pipeline, the regulatory approval, the integration with existing workflows, and the maintenance over time. The techniques from NLP, CV, RL, and MLOps are the building blocks; this domain shows what gets built with them.