Engineering AI Systems - A Summary of Architecture and DevOps Essentials

Engineering AI Systems
Architecture and DevOps Essentials
Prof. Dr. Ingo Weber
Chair of Information System Development and Operation
Technical University of Munich, School of CIT, CS Department
Fraunhofer-Gesellschaft, Director for AI & Innovation
Coauthors:
Len Bass
Q.Lu, L. Zhu
Book website:
https://research.csiro.au/ss/
team/se4ai/ai-engineering/
Unless otherwise specified, figures are from the book and subject to copyright

Introduction
AI & Foundation Models
AI System Life Cycle
Qualities
Summary
2
Engineering AI Systems | Prof. Dr. Ingo Weber

2024 study:
• 80% of AI projects do not go into production
− https://www.rand.org/pubs/research_reports/RRA2680-1.html (2024)
− Included ML projects but not projects that used pre-trained LLMs (prompt
engineering)
− This is twice the rate for non-AI projects
−or, from the opposite viewpoint: the success rate is only a third of that of
regular software projects (20% vs. 60%)
Optimism about AI systems persists
3

Quality of software: how well does it do its job – along all relevant dimensions
→ Does the software fulfil the (implicit and explicit) requirements in terms of functionality and quality attributes?
ISO/IEC 25010:2011 Quality Model
Software quality
4
AI systems with high quality: meet quality
expectations in terms of conventional software
portions, AI model(s), and the combination

DevOps is a set of practices intended to reduce the time between committing a
change to a system and the change being placed into normal production, while
ensuring high quality. [DevOps book, 2015]
AI engineering is the application of software engineering principles and
techniques to the design, development, and operation of AI systems.
[AI Engineering book, 2025]
The term MLOps, analogous to DevOps, encompasses the processes and tools not
only for managing and deploying AI and ML models, but also for cleaning,
organizing, and efficiently handling data throughout the model life cycle.
[AI Engineering book, 2025]
Terminology
5

Introduction
Qualities
Summary
6

The specifics of the system architecture depend on the type of AI being used, nowadays mostly:
• Narrow ML Models or
• Foundation Models (or combinations)
Types of AI systems
Engineering AI Systems | Prof. Dr. Ingo Weber 7

Overview of Model Types
▪ Narrow ML Models – Designed for specific tasks using custom-collected and labeled datasets.
▪ Foundation Models (FMs) – Large, pretrained on extensive datasets for general purposes, adaptable to a variety of
specific tasks.
• Large language models (LLMs) are a subtype of FMs
▪ Symbolic AI (a.k.a. Good-old-fashioned AI, GOFAI): rule-based systems, symbolic reasoning, AI planning,
knowledge graphs, etc.
• Not discussed in depth here
Terminology: Narrow ML Models and Foundation Models
8
Artificial Intelligence
Symbolic AI
Machine Learning
Narrow ML Foundation Models
Large Language Models

… are statistical models
Intended for a specific task
Input to a narrow ML model is a data item consisting of a set values of labelled independent variables.
Output is a generated value of a dependent value.
What are narrow ML models used for?
Suitable for three main purposes
• Classification – assigns a category to an input. e.g. this email is spam
• Regression – returns a continuous value to an input. E.g.,
− this particular process will take 3 days to complete,
− this pipe will break in the next 6-24 months (and should be replaced)
• Clustering – groups similar items. E.g. this data set has 10 groups, clustered by age
Narrow ML models

Ethical concerns
Interpretability and explainability
Generalization and overfitting
Robustness and adversarial attacks
What are the problems with narrow models?

What are Foundation Models?
A Foundation Model (FM) is trained on an
extensive and diverse dataset, often comprising
billions or even trillions of data points.
The training data is largely unlabeled,
FMs are general purpose but can be customized
for particular applications.
Large language models (LLMs) are a type of
FMs.
AI-generated figure

Natural Language Processing
Text Machine translation
Question answering
Summarization
Image generation and classification
Object detection
Code Generation
Video / Audio generation
Brainstorming
Rapid prototyping
…and many other applications
What are Foundation Models used for?

Data Privacy and Security
Sensitive Data Leakage / Exposure
Misuse and Misinformation
Deepfakes and Fake Content
Hallucination? Confabulation?
What are some problems with Foundation Models?
13

What are some problems with Foundation Models?
Hallucination vs. Confabulation
14
„Create a photo of the moon surface“
(not an actual example) → Hallucination
…vs. confabulation
AI-generated figures

▪ Recap: Compact Overview of Model Types
• Narrow ML Models - Designed for specific tasks using custom-collected and labeled
datasets
• Foundation Models (FMs) - pretrained for general purposes, i.e.: not one specific
purpose, adaptable to a variety of specific tasks
▪ Current Use in AI Systems
• Most AI systems from before 2024 integrate narrow ML models and non-AI components
that handle specialized tasks and system functionalities
• FMs are being increasingly explored and utilized for their adaptability and broad
application potential
Central design decision in AI system:
Choosing between Narrow ML Models and FMs
15

▪ Advantages of Narrow ML Models
• Tailored to specific tasks, potentially offering higher precision for particular applications
• Greater control over the model development process, allowing for customized
adjustments and integration of safety measures (guardrails)
▪ Challenges with Narrow ML Models
• High cost of data collection, labeling, and computation during training
• Potentially limited by the quality and quantity of available training data
Advantages / Challenges with Narrow ML Models
16

▪ Advantages of Foundation Models:
• Can reduce the amount of human labor and time required for model development.
• Offer a flexible base that can be fine-tuned for accuracy improvements across diverse
tasks.
• Economical in terms of reuse and scalability, especially when shared within an
organization or accessed via API.
▪ Challenges of FMs:
• Training effort infeasible for many organizations
• Potentially high cost / compute requirements for inference (i.e. runtime use)
• Updating
• Limited grounding
• Hallucination / confabulation
Benefits / challenges of using Foundation Models
17

▪ Choosing between model types depends on specific project needs, resource availability,
and desired control over the model characteristics
▪ Organizations must weigh the ease of implementation and broad applicability of FMs
against the specificity and customizable aspects of narrow ML models
Trade-offs and Decision Factors
18

▪ Cost is a key factor in deciding between Narrow ML models and FMs
▪ Cost factors include:
• Development cost – similar for system, lower for using existing FMs
• Maintenance cost – depends on customization / model size
• Operation costs – typically higher for FMs
Costs of Narrow ML Models and FMs
19
AI-generated figure

Narrow ML Models can Co-exist with FMs
20

Narrow ML Models can Co-exist with FMs
21

▪ Adapting foundation models (FMs) to specific applications involves various strategies to refine
their functionality and ensure alignment with use case requirements.
▪ Techniques for Customizing FMs
• Prompt Engineering - Crafting specific prompts or input sequences that guide the FM towards
producing desired outputs.
• Retrieval-Augmented Generation (RAG) - Enhancing the FM's responses by dynamically
incorporating external knowledge or data during the generation process.
• Fine-Tuning - Adjusting the model’s parameters on a specific dataset to improve performance
for particular tasks.
• Distilling - Simplifying and compressing the model to enhance efficiency while retaining
performance, useful for deployment on resource-constrained environments.
Customizing Foundation Models
22

Customizing Foundation Models
23

Prompt engineering is the easiest for organizations
Implementing RAG is more difficult
Fine tuning and distilling require sophistication
Creating a bespoke FM requires a high degree of sophistication
Complexity
Complexity vs time/organizational maturity

▪ Guardrails are designed to monitor and control the
inputs and outputs of
• foundation models
• users
• RAG
• external tools
▪ Goal is to meet specific requirements, including
• function
• accuracy
• aspects needed due to policy (AI Act, etc.),
standards and laws.
Guardrails
25
Stock image, CC licence

▪ Input guardrails are applied to the inputs received from users, and their possible effects include
refusing or modifying user prompts.
▪ Output guardrails focus on the output generated by the foundation model, and may modify the
output of the foundation model or prevent certain outputs from being returned to the user.
▪ RAG guardrails are used to ensure the retrieved data is appropriate, either by validating or
modifying the retrieved data where needed.
▪ Execution guardrails ensure that the called tools or models do not have any known
vulnerabilities and the actions only run on the intended target environment and do not have
negative side-effects.
▪ Intermediate guardrails can be used to assert that each intermediate step meets the necessary
criteria.
Different Types of Guardrails
26

Types of guardrails in NVIDIA NeMo
27
“NeMo Guardrails is an open-source toolkit for easily adding programmable
guardrails to LLM-based conversational applications.”
Slight changes in terminology: guardrails → rails; RAG guardrails → retrieval rails;
Intermediate guardrails → dialog rails.
Source: https://github.com/NVIDIA/NeMo-Guardrails

Introduction
Qualities
Summary
28

Quality is determined by:
• Software architecture
• Development processes
• Code quality
• Model quality
Quality Influences in AI Systems
System
qualities
Development
processes

30
System Build
Create an executable
artifact
System Release
Approve for deployment
System Test
Ensure high test coverage
& automate tests as much
as possible
Design
Design architecture to fulfill
requirements, achieve goals, and
support other activities
System Dev
Perform normal development
activities & create scripts for
other activities
Deploy
Move into production environment
Operate & Monitor
Execute system,
gather measurements,
display metrics,
detect anomalies
Analyze
Analyze measurements taken
during operation, user feedback,
new developments, etc.
Model Build
Create an executable artifact
Model Release
Approve for integration in system
Model Test
Test accuracy, risks, biases, etc.
AI Model Dev
Model selection, exploration,
hyperparameter tuning, data
management, model preparation
and training, model evaluation
AI
Model
Dev
Deploy
Operate &
Monitor
Build
Release
Analyze
Test
Build
Dev
Rel.
Test

AI models are just one component (each)
The AI model processes are for developing and testing the AI model
32
AI
Model
Dev
Deploy
Operate &
Monitor
Build
Release
Analyze
Test
Build
Dev
Rel.
Test

Sufficient training data – 10x number of parameters in the model
Distribution of training data
• Reflect distribution of requests
• Corrected for biases
• Adjusted to reflect security concerns
Data quality is important

Data drift – training data is not representative of actual data after some period of time. E.g. housing prices
have changed
Environmental change – e.g. training data represents housing prices in urban areas, attempt to use model
for rural areas.
Regulatory changes – ongoing regulations can affect issues such as fairness, transparency, accountability,
and human oversight
- Imagine the road rules changing…
Problems with data

AI models are just one component (each)
The AI model processes are for developing and testing the AI model
The remaining processes are for both the AI and non-AI components
• Create non-AI components
• Build the AI system with both AI components and non-AI components
• Test it, deployment it, and operate it
35
AI
Model
Dev
Deploy
Operate &
Monitor
Build
Release
Analyze
Test
Build
Dev
Rel.
Test

▪ Interdisciplinary teams should be set up to jointly develop the portion of the design that affects
both the AI and the non-AI modules
▪ Interdisciplinary teams can frequently have problems with vocabulary and cultural differences
▪ Bruce Tuckman characterized team formation as
“forming, storming, norming, performing”
▪ Different expertise may be needed at different times, e.g., UX, security, …
Interdisciplinary Teams
36

Co-design and development
• Allows negotiation about how functions should be divided
• Ideally, some developers will have expertise across both fields
Using microservices
• Some components may not make good microservices, e.g. due to communication issues (latency, etc)
Design for modifiability
• Intermediary between the client of a component and the component itself, e.g. translating clients’
invocations into the new model interface
Design

Development of non-AI modules is similar to non-AI systems; testing (and developing tests) is not
A module is a coherent piece of functionality
A service is constructed through the integration of multiple modules and their supporting libraries during the build step.
• Using the language and technologies the team chooses
Unit test:
• Tests each module with defined inputs and expected outputs
• During development and again during the testing phase
• Unit tests cover
− Common cases: typical scenarios
− Edge cases: boundary of input ranges or special conditions
− Negative cases: invalid inputs where the module should fail gracefully
Develop Non-AI Modules

CI triggered when releasing a new AI model or committing a
change
CI retrieves all elements, build a new image, run tests
If tests pass, the image is deployed into a staging server
CI server also creates build meta data, such as version
manifest, dependency graph, package manifest
Build

Repeat model/unit tests from AI/non-AI module development
System-wide test for end-to-end execution of the system
• User stories
• Regression testing
• Compatibility testing
• Efficiency/performance testing
• Security testing
• Compliance testing
Non-repeatable testing results
• Test database has not been restored
• Have background tasks
• Probabilistic AI models may produce different results at different times
→ need to control all sources of randomness to achieve repeatable tests
Test

Have a single, well-defined release and deployment process – unchanged DevOps principle
• May require a human to release the system to deployment for regulatory/legal reason
− Frequency of new releases: occur on a defined schedule or continuously
• Continuous release/deployment patterns
− Blue/green deployment: deploy multiple instances of a new service simultaneously and route all new requests to
instances of the new service while the instances of the old service are gradually drained and terminated
− Rolling upgrade: gradually replace instances of old service with new service, typically one at a time
− Canary testing: release a new version to a selected group of users to test the new version
− A/B testing: compare two versions and determine which one is better
− Roll back: revert to a previous version
Release and Deploy (Cont’d)

Additional reasons for updating AI systems
• Data drift
− Data distribution the model encounters in production can be different from the data it was originally trained
• Improved data
− The model may encounter additional/updated data in production that wasn’t available during the initial training
• Changing requirements
− The requirements for the model may change over time, e.g., add new features to the model or improve the
accuracy
• Biased/vulnerable models
Release and Deploy

Collect and analyze data to inform the next round of development
Monitoring Sources and metrics
• Infrastructure Metrics
− resource utilization including CPU usage, memory consumption, I/O, and network traffic of VMs and
containers
• System code metrics
− Application-specific metrics such as number of active users, number of requests created, API response
time
− Need higher-level metrics to assess AI systems, e.g., sentiment, precision of answers, number of
retry prompts
• Logs
− Text records generated by services and stored locally or moved to databases for analysis, such events,
errors, state changes
Operate, Monitor, and Analyze

Introduction
Qualities
Summary
44

A system is constructed to satisfy organizational objectives, which can
be manifested as a set quality requirements.
Quality requirements for a system are formalized as quality attributes
(QAs)
• A QA is a measurable or testable property of a system that is used to
indicate how well the system satisfies the needs of its stakeholders beyond
the basic function of the system
• Stated for both AI and non-AI portions
Qualities

Reliability
• Operate over time under various specified conditions and without failure.
• AI model performs effectively when introduced to new data
− This new data should be within or similar to the distribution of the training data, not out of distribution (OOD).
Robustness
• Operate despite unexpected inputs
• AI model is able to handle out-of-distribution (OOD) data and adversarial input
Resilience
• Operate despite changes in the environment
• In AI, this is characterized by the model’s intrinsic ability to continue functioning effectively amidst significant disruptions.
Reliability, Robustness, Resilience

Performance

Efficiency: utilize resources effectively while maintaining acceptable speed and
responsiveness, in addition to fulfilling its functional requirements.
• Throughput: number of tasks processed within a specific time frame
• Latency: time it takes for a system to complete a task
• Scalability: handle increasing workloads without degrading performance
− Training scalability: train AI models as training task size increases (e.g. number of
model parameters, data volumes)
− Inference scalability: deploy and operate AI systems for inference as the model size
or the number of requests increases
• Resource utilization: how efficiently a system uses its computational resources
− Training time resource utilization: the effective use of computational resources for
training
− Inference time resource utilization: the model’s resource efficiency in making
predictions or generating outputs
Efficiency

Architecture Design
• Use an interface style like GraphQL
• Optimal placement of containers and VMs
• Caching
• Edge computing
• Mode switcher
− Selecting components/models
− FM as a software connector
Architecture Approaches to Improving Efficiency

Model Choice
• Assess resource required for both training and inference
• Compared to deep neural networks, simpler models like decision trees or naive Bayes are more resource friendly for
both training and execution.
Hyperparameters
• Adjusting hyperparameters (e.g. learning rate) to reduce the number of training iterations
• Employing regularization hyperparameters (like L1 or L2 penalties) to simplify the model
• Utilizing early stopping involves halting the training process
• Reducing the precision of the model parameters
Data sampling
• Using random sampling to select a representative sample of dataset for training
Feature engineering
• e.g. reducing the number of features or using hashing to handle high cardinality features
Operation
• Process multiple data points simultaneously during inference
• Use lazy loading or model checkpointing to load only the necessary parts of the model at runtime
Process Approaches to Improving Efficiency

Accuracy refers to a system’s ability to produce expected or true results.
• AI: how well the AI system’s predictions or outputs align with the expected or true results.
Accuracy Metrics for Classification Models
• Accuracy Rate
• Precision
• Recall
• F1 Score
Accuracy Metrics for Regression Models
• Mean Squared Error (MSE)
• Root Mean Squared Error (RMSE)
• Mean Absolute Error (MAE)
• R-squared (R2)
Accuracy Metrics for Foundation Models
• Bilingual Evaluation Understudy (BLEU) Score
• Recall-Oriented Understudy for Gisting Evaluation (ROUGE) Score
Accuracy

Repeatability: produce the same output for the same input under identical conditions.
Some systems may not be deterministic because of factors like multi-thread design or other background activities.
AI systems maybe non-deterministic
• Rely on randomness during training, such as initializing weights or selecting optimization paths.
Output variability is desirable in certain AI applications, like creative writing or generating various strategic
recommendations.
Performance measurement, such as with benchmarks, requires multiple executions and statistical analysis.
Repeatability

Architecture design
• Multi-model decision-making
− Using different models to perform the same task or make a single decision.
− Consensus protocols can be defined to make a final decision, such as taking the majority decision or accepting only
unanimous results from all models.
− The end user or operator reviews the output from multiple models and makes the final decision.
Approaches to Improving Accuracy

Choice of hyperparameters
• Learning rate
− Too low rate: slow progress
− Too high rate: diverging training - disrupt the learning process
• Number of trees in a random forest
− too many trees may cause overfitting
− too few may lead to underfitting

Data preparation
• Sufficient and representative training data
• Bias-free data
• Incorporating samples of OOD data - handle unexpected inputs
• Addressing Missing Data
− Use imputation techniques to estimate and fill in missing values or remove rows/columns that contain excessive
missingness
• Managing Outliers
− Clip outliers to a certain range or eliminate them entirely
• Splitting the data into three main sets with same distribution (80%, 10%, 10%)
− Use cross-validation techniques to split the data if the training dataset is too small

Model preparation - feature engineering
• Scaling features to a common range so they contribute equally to training
− e.g. using standardization or normalization technique for larger range or high
value features
• Converting categorical data to numerical values so model can understand the data
− e.g. using one-hot encoding or label encoding
• Deriving new features from existing data
− e.g. calculating "time since last purchase"
• Applying transformations like log or square root to make the relationship between
features and target variables more linear
• Selecting the most influential features for the model
− e.g., using chi-square tests for categorical targets or deriving feature importance
scores from ensemble methods like random forests.
• Reducing the number of features using techniques like Principal Component
Analysis (PCA)

Model generation - model evaluation provides insights into how well the model is likely to perform on unseen data
• Evaluation metrics: Evaluate them once the model is full trained and monitor them through the training phase
− Metrics for classification models: accuracy rate, precision, recall, F1 score
− Metrics for regression models: Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Mean Absolute
Error (MAE), R-squared (R2)

Operation
• The accuracy of a model is assessed based on known ground truth values.
• It's essential to periodically test the model for accuracy to ensure that it remains effective over time.
− If accuracy issues are detected, the model may need to be retrained with new data or adjusted to address any
identified issues.
− The updated model should undergo a rigorous deployment process to ensure that it meets the operational
standards and accuracy requirements before being fully implemented.
• Accuracy Testing Techniques
− Confusion Matrix - visualizing the performance of a model across different classes - helps identify imbalances
− Statistical tests like the Kolmogorov-Smirnov test can be used to compare the distribution of new data against the
training data – detecting data/concept drift

FM-based Systems: general-purpose nature makes it hard to achieve consistent accuracy across various domains
without specific adjustments
• Fine-tuning: Adapting the model on a more specific dataset
• Retrieval-Augmented Generation (RAG) - Integrating external data during generation
• Prompt Engineering - Crafting inputs (prompts) that guide the model to generate more accurate and contextually
appropriate outputs.
• Reinforcement Learning from Human Feedback (RLHF) - Adjusting the model's parameters based on feedback from
human evaluators.
• Adversarial Testing and Benchmarks - Testing the model against carefully crafted adversarial inputs or established
benchmarks to ensure robustness and accuracy.
• Acceptance Testing and Test Case Assessment - Systematically testing the model/system to verify performance
before full-scale deployment.

CIA Properties
• Confidentiality – only authorized parties can access data or resources.
• Integrity - data is accurate, complete, and has not been altered or modified by unauthorized parties
• Availability – there is timely and reliable access to data, systems, and services.
Authentication means that the persons or systems trying to access a service are who they claim to be.
Authorization refers to the rights the user has, i.e., they are authorized to perform certain operations like accessing a file or
approving an invoice.
Security

Attacks in an AI System

Architectural Approaches – based on zero trust
• Authorization, encryption, least functionality, limiting access, privilege minimization
Process Approaches
• Threat modeling and risk assessment for identifying/assessing vulnerabilities
• New attacks to AI/FM systems: use AI to track and analyze attacks
• Maintaining versions of data items as well as their lineage to ensure the integrity of the model
• Adversarial testing to detect vulnerabilities
• Logging and auditing can help to detect anomalies
Approaches to Mitigating Security Concerns

More complex
• Not only data access and storage but also the outputs and decisions of the AI systems which can potentially reveal
sensitive information
Organizational approaches: appoint executives such as Chief Privacy Officers (CPOs) or Data Protection Officers
(DPOs). Their responsibilities include
• Developing and enforcing data privacy policies
• Monitoring organizational compliance with privacy laws
• Handling data subject requests e.g. access or deletion
• Managing data breaches
• Educating and training employees on data privacy best practices
Process and architecture: Implementing LOCKED rights—Limit, Opt-out, Correct, Know, Equal, Delete
• E.g., machine unlearning, guardrails
Privacy in AI Systems

Fairness is not only about the outcomes of an AI system but also about the decision processes and model development
processes, including data collection, algorithm design, ongoing management of the system.
• For instance, in recruitment AI tools, fairness is challenged by historical biases present in the training data. An AI
system might inadvertently learn to prefer candidates from a certain demographic background, etc.
Organizational strategies
• Fairness teams: comprising members from data science, engineering, legal, and ethics departments to identify
fairness concerns at various stages.
• Third-Party Expertise: auditing their AI systems. These experts provide impartial assessments and recommend
strategies to address potential biases and ensure fairness.
• Employee Training: increasing awareness about the impact of bias and the importance of fairness.
Architecture
• Guardrails and monitor
• Logging
Process Approaches
• Fairness Metrics: Demographic Parity, Equal Opportunity, Equalized Odds, Predictive Parity, Treatment Equality
• Fairness tools: AI Fairness 360 by IBM, Fairlearn by Microsoft, PAIR Tools by Google
Fairness

Monitorability: ability to monitor and track predefined AI system/model quality metrics
• More reactive: identifying issues as they occur
• Solely relying on predefined metrics
Observability: ability to gain insights into the inner workings of the AI system at system level, the model level, and the
model development and deployment pipeline level, and into its operating infrastructure.
• More proactive: understanding an AI system, detecting data, model and concept drift early, and preventing incidents
• Leverages broader data sources, including metrics, logs, traces, events generated by the AI system
• Tools: Repositories for code and data, system logs, resource utilization monitoring infrastructure
Monitorability and Observability

Maintain logs to record system activity, ensuring all relevant events are captured for later analysis.
• Data Lineage Tools: track the history and flow of data used in model predictions
• Software Bill of Materials (SBOM): provide a detailed lineage of non-AI components
Explainable AI techniques
• Local explanation techniques : instance-based explanations for understanding the feature importance and correlations
that led to the specific outputs, e.g. LIME and SHAP
• Global explanation techniques: understand the general behavior of an AI system by using a set of data instances to
produce explanations
− Visualizing the relationship between the input features and model’s output over a range of values
− Global surrogate models such as tree-based models or rule-based models
• Foundation model-based systems: think aloud vs. think silently
Co-versioning registry
Independent overseeing agents
Approaches to Ensure Observability

Introduction
Qualities
Summary
67

Notes for instructors:
• Slides for the whole book are available
• All chapters include discussion questions
68
Book chapter overview
1. Introduction
2. Software Engineering Background
3. AI Background
4. Foundation Models
5. AI Model Lifecycle
6. System Lifecycle
7. Reliability
8. Performance
9. Security
10.Privacy and Fairness
11.Observability
12.Case Study: Using a Pretrained Language Model for Tendering
13.Case Study: Chatbots for Small and Medium-Sized Australian Enterprises
14.Case Study: Predicting Customer Churn in Banks
15.The Future of AI Engineering

Summary
69

Engineering AI Systems
Architecture and DevOps Essentials
Prof. Dr. Ingo Weber
Book website:
https://research.csiro.au/ss/
team/se4ai/ai-engineering/

Engineering AI Systems - A Summary of Architecture and DevOps Essentials

More Related Content

Similar to Engineering AI Systems - A Summary of Architecture and DevOps Essentials

More from Ingo Weber

Recently uploaded

Engineering AI Systems - A Summary of Architecture and DevOps Essentials