AI Engineer

The AI Engineer knowledge library

Topics in AI engineering

Connected chapters on AI systems, applications, and organizations, with the conference talks behind them.

60 chapters · 2281 talk selections

Browse by subject

Foundations

Software Engineering Fundamentals

How contracts, boundaries, state protection, testing, debugging, and controlled releases make applications dependable even when model behavior is uncertain.

42 talks · 50 speakers

Machine Learning Fundamentals

How examples, objectives, fitted parameters, and independent evidence turn data into useful predictions—and why prediction alone does not tell you what action will work.

46 talks · 55 speakers

Tokenization

How text, message structure, and special markers become model-specific vocabulary IDs—and how that representation governs boundaries, reconstruction, request limits, and charges.

9 talks · 7 speakers

Embeddings and Representation Learning

How learned vectors preserve task-specific distinctions, acquire geometry, support larger systems, and fail when similarity is mistaken for truth.

28 talks · 32 speakers

Transformers and Attention

How sequences become contextual states, how those states produce predictions, and what information access can—and cannot—guarantee.

18 talks · 18 speakers

Prompting and In-Context Learning

How instructions and demonstrations guide a fixed model—and how to test whether the resulting behavior transfers beyond the prompt-development cases.

32 talks · 39 speakers

Learning and Reasoning

Pretraining and Midtraining

How objectives, data exposure, schedules, compute, and checkpoint lineage establish and extend a model’s capabilities.

31 talks · 35 speakers

Post-training and Alignment

How demonstrations, preferences, and rewards change a pretrained model—and how to determine what the changed checkpoint actually does.

54 talks · 56 speakers

Reasoning and Test-Time Compute

How intermediate work, alternative solutions, and checking improve answers—and how to decide when that work is worth doing.

23 talks · 24 speakers

Synthetic Data

Generated examples help only when they add verified coverage of a defined capability gap and improve performance on evidence kept independent from their creation.

40 talks · 45 speakers

RL Environments and Simulators

Build executable tasks whose observations, actions, rewards, resets, and recorded outcomes mean what you intend.

33 talks · 37 speakers

Continual Learning

Acquire new capabilities through successive updates while preserving useful behavior—and the ability to keep learning.

11 talks · 11 speakers

Distillation

Train a model from another model’s behavior—and determine what actually transferred.

22 talks · 25 speakers

Knowledge, Context, and Connectivity

Data Quality and Curation

How to choose, transform, label, release, and maintain data whose meaning remains fit for an AI system’s intended use.

26 talks · 30 speakers

Retrieval-Augmented Generation

Turn external information into useful answers without losing the conditions, sources, and limits that make those answers defensible.

51 talks · 61 speakers

Search and Retrieval

How search systems acquire an eligible collection, turn it into candidates, rank useful results, and serve them under permission, freshness, coverage, and latency constraints.

41 talks · 48 speakers

Knowledge Graphs

When explicit identities, typed relationships, and connected evidence justify their construction—and how to keep every graph assertion sourced, revisable, and useful.

26 talks · 26 speakers

Context Engineering

How to construct the smallest sufficient, current, permitted, and recoverable input for each model call.

82 talks · 97 speakers

Agent Memory

Retain useful experience across interactions without turning old, mistaken, or unauthorized information into a lasting premise.

51 talks · 55 speakers

Model Context Protocol

Reusable capability exchange, from discovery and authorization to execution and uncertain outcomes.

50 talks · 54 speakers

Agents and Software Development

Agent Engineering

Choose useful actions from changing observations, keep them within delegated limits, and establish what the resulting work accomplished.

93 talks · 99 speakers

Software Factories

Coordinate automated software production without losing control of requirements, acceptance, release, or unfinished work.

44 talks · 50 speakers

Coding Agents

Turn software requests into inspected, tested, and integrated changes—not just plausible code.

77 talks · 85 speakers

Computer Use

Observe graphical state, identify the intended control, deliver the right input, and verify the authorized outcome.

45 talks · 52 speakers

Personal Agents

Continuing assistance that remembers what matters, carries work forward, and stays within the person's control.

30 talks · 32 speakers

Multi-Agent Systems

Divide work and discretion without losing task meaning, responsibility, or control of the complete outcome.

33 talks · 43 speakers

Inference and Infrastructure

Inference Engineering

How a serialized model request becomes a streamed completion—and how state, scheduling, memory, and workload shape determine latency, throughput, and sustainable capacity.

46 talks · 57 speakers

Quantization

How numerical mappings, calibration and execution support turn fewer bits into useful savings—and where those savings can fail.

18 talks · 24 speakers

Local and On-Device AI

Choose where inference belongs, fit the complete workload to the device, and preserve clear promises about readiness, connectivity, and data.

67 talks · 84 speakers

Sandboxes and Execution Isolation

How to run generated or otherwise untrusted code with bounded files, networks, authority, resources, lifetime, and tenant exposure—without pretending teardown reverses effects outside the boundary.

35 talks · 36 speakers

Model Routing and LLM Gateways

Choose suitable models and serving paths while preserving permissions, request meaning, resource limits, and honest failure behavior.

26 talks · 33 speakers

AI Platform Engineering

Build shared capabilities that make AI systems easier to deliver and operate while keeping authority, running versions, capacity, and responsibility visible.

31 talks · 39 speakers

AI Cost and Performance Engineering

Reduce the resources needed for useful work without losing the quality, responsiveness, or operating limits that make the system worth using.

35 talks · 43 speakers

Evaluation, Reliability, and Trust

Evals and Benchmarks

How to turn intended work into bounded, reproducible evidence—and know exactly where that evidence stops.

78 talks · 89 speakers

Observability

Design operational evidence that can reconstruct an AI execution, locate consequential failures, attribute time and cost, and state honestly what remains unknown.

67 talks · 81 speakers

AI Security

Trace attacker influence through data, models, tools, execution, and external effects—and place enforceable controls where that influence would otherwise gain authority.

60 talks · 70 speakers

Privacy and Data Governance

Turn decisions about appropriate information use into enforceable limits on collection, access, disclosure, modification, retention, and deletion.

28 talks · 37 speakers

Multimodal and Physical AI

Multimodal Models and Applications

How to combine text, images, audio, video, and other signals without losing the relationships that make their conclusions meaningful.

32 talks · 42 speakers

Computer Vision

How images and video become predictions about objects, regions, space and events—and how to choose representations that preserve the information a task needs.

23 talks · 28 speakers

Voice and Real-Time AI

How speech models, audio transport, and conversational policies cooperate to listen, respond, stop, and recover.

32 talks · 42 speakers

Generative Media

How images, sound and video are synthesized, directed, revised and assessed.

38 talks · 42 speakers

World Models

Learning how environments change—and establishing when predicted futures are useful for planning, training, and simulation.

14 talks · 16 speakers

Robotics

Connecting goals, measurements, motion, and learning in a physical system that must keep responding while the world changes.

13 talks · 15 speakers

Products and Organizations

Design Engineering and AI Interfaces

How to help people understand, direct, review, interrupt, and recover from AI behavior that may be uncertain, delayed, or partly wrong.

54 talks · 63 speakers

Workflow Automation

Redesign complete processes, allocate decisions deliberately, and keep people, records and external effects under clear control.

50 talks · 52 speakers

Forward Deployed Engineering

Turn customer work into a useful, maintainable system—and turn deployment findings into better product decisions.

51 talks · 58 speakers

Enterprise AI

Adopt AI by changing how work is delivered, who can decide, and how useful outcomes are sustained.

72 talks · 81 speakers

Agentic Commerce

Delegate commercial choices while preserving valid terms, purchasing authority, and responsibility for what remains owed.

26 talks · 28 speakers

AI in Sales and Marketing

Improve customer understanding, communication, and follow-through—and distinguish useful commercial work from increased activity.

26 talks · 30 speakers

AI Engineering Leadership

Choose worthwhile investments, fund the capabilities they require, and keep authority aligned with consequences.

24 talks · 30 speakers

Applied Systems and Discovery

Recommendation Systems

Select useful items, learn from incomplete feedback, and measure whether the resulting policy improves the product.

17 talks · 23 speakers

AI in Finance

Build useful financial assistance by preserving meaning, separating analysis from authority, and checking the complete workflow.

44 talks · 51 speakers

AI in Healthcare

Build useful assistance around clinical work, meaningful records, accountable people, and evidence of completed care.

34 talks · 39 speakers

More from the talk archive 20 collections

Explore additional talk collections. These archive topics are separate from the chapters above.