YÖK Data Analysis School 2025-2026: Artificial Intelligence Training

I attended 13 common classes at the YÖK Data Analysis School between 8 October 2025 and 9 January 2026, followed by 23 sessions in the Artificial Intelligence and Machine Learning and Artificial Intelligence and Facilitating Tools modules. The training period concluded on 12 May 2026.

The program was organized under the Council of Higher Education of Türkiye (YÖK), coordinated by the Marmara University Institute of Population and Social Research, with academic contributions from Middle East Technical University, Istanbul Technical University and Boğaziçi University. The common classes covered data preparation, Excel, SPSS, statistics and Python. The two modules focused on machine learning, generative AI and AI-assisted research and analysis tools.

Program Structure and Selected Modules

The program comprised a common stage followed by specialization. The 13 common classes covered measurement scales, data cleaning, statistical inference, Excel, SPSS, Python, NumPy and pandas. After selecting my specialization, I attended 23 sessions across two AI modules between 6 February and 12 May 2026. The dated session notes appear below in chronological order.

Artificial Intelligence and Machine Learning

The Machine Learning module covered supervised, unsupervised and reinforcement learning, including classification, regression, clustering, dimensionality reduction and model evaluation. The algorithms and methods included K-means, Naive Bayes, decision trees, random forests, support vector machines (SVMs), principal component analysis (PCA) and linear discriminant analysis (LDA). The classes also addressed train-validation-test separation, data leakage, overfitting, regularization and performance metrics.

Deep learning topics included neural networks, convolutional neural networks (CNNs), attention and transformer architectures. Subsequent sessions addressed language models, tokenization, generative models, quantization, parameter-efficient fine-tuning, retrieval-augmented generation (RAG), tool-using agents and local inference. Related Academy notes: Statistical Learning and Machine Learning, Artificial Neural Networks, Large Language Models and Retrieval-Augmented Generation (RAG).

Artificial Intelligence and Facilitating Tools

The Facilitating Tools module addressed language-model-assisted text classification, qualitative coding, cleaning academic publication records, literature search, source verification and reporting. Other subjects included visualization with R and Python, automatic transcription, model APIs, presentation tools and workflow automation.

The sessions examined prompt formulation, comparisons on consistent datasets, human review of automated labels, privacy in data transfers and research methods based on verifiable sources. Related Academy notes: Large Language Models, Natural Language Processing and Prompt Engineering.

Participation and Certification

I completed the attendance and assessment requirements and passed the examinations for both selected modules. I received one participation certificate and one achievement certificate for each module, four documents in total:

• Artificial Intelligence and Facilitating Tools — Participation Certificate
• Artificial Intelligence and Facilitating Tools — Achievement Certificate
• Artificial Intelligence and Machine Learning — Participation Certificate
• Artificial Intelligence and Machine Learning — Achievement Certificate

The certificates contain QR codes for verification. All four documents relate to the two AI modules completed during the 2025–2026 program.

Data Analysis School 2025–2026 training modules

Boğaziçi University — Data Analysis School announcement

Common Classes — Data Analysis, Statistics and Python (2025–2026)

The 13 common classes held between 8 October 2025 and 9 January 2026 covered data literacy, observational units, variables, measurement scales, Excel, SPSS, descriptive statistics, sampling and hypothesis testing. The final four sessions focused on Python programming, numerical computation with NumPy and data preparation with pandas. The dated notes follow.

8 October 2025 — Foundations of data analysis, statistics and computational thinking

The opening class covered statistics as both a method of describing observations and a framework for inference. Data science was described in terms of statistical methods, computing and domain knowledge. Observational units, variables and measurement were distinguished through examples of numerical and categorical data. Storing a category as a numeric code does not make the underlying variable quantitative; its measurement scale determines which summaries and analyses are appropriate.

Binary representation, basic logic and levels of programming-language abstraction were introduced. Data storage, transformation and computation connected statistical reasoning with programmatic processing. Computing resources, dataset size and data quality were discussed as practical factors in contemporary analytics and machine learning.

Related Academy notes: Probability and Statistics

15 October 2025 — Observations, variables and data preparation in Excel

The tabular data model was introduced through observations represented by rows and variables represented by columns. Unique record identifiers, the joining of datasets, and the distinction between microdata and aggregates were examined. Consistent variable names, explicit data types and the representation of missing observations were presented as prerequisites for defensible analysis. Treating a missing value as zero was examined as a potential source of material changes in calculations and conclusions.

Nominal, ordinal, interval and ratio scales were compared. It was noted that numeric category codes did not, by themselves, imply an ordering or meaningful arithmetic operations. Excel formula references, conditional operations, validation and category coding were discussed as practical data-preparation techniques. The session established that the data dictionary and the level of measurement must be resolved before selecting analytical procedures.

22 October 2025 — Data transformation, summarization and visualization in Excel

Practical spreadsheet operations included selecting and filtering data, freezing headings, hiding cells and understanding how references change when formulas are copied. Relative, absolute and mixed references were compared, particularly where a calculation is applied consistently across a large number of observations. Transposition, text and date functions, lookup operations and PivotTables were used to explain common cleaning and restructuring workflows.

Chart selection was considered in relation to the variable type and the analytical question. Axes, units and category labels were discussed as factors affecting the interpretation of a visual result. Versioning working files and retaining untouched source data were treated as requirements for a reversible analysis. Summaries and charts were distinguished from additional observations: they re-express existing records at another level of aggregation rather than create independent evidence.

5 November 2025 — SPSS data structures and variable metadata

SPSS Data View and Variable View were used to distinguish observed values from the metadata that describes their meaning. Data (.sav), output (.spv) and syntax (.sps) files were compared in terms of reproducibility. Numeric, string and date types were separated from nominal, ordinal and scale measurement levels. Value labels were considered a representation layer, not a substitute for determining the meaning of a variable.

The session also addressed missing-value codes, weighting and scripted operations. A survey dataset was used to illustrate variable documentation, respondent confidentiality and the role of survey weights in interpretation. Reading the codebook and checking the correspondence between metadata and actual values were identified as necessary steps before using frequency tables or inferential procedures.

12 November 2025 — Descriptive analysis and transformations in SPSS

A social research dataset provided the context for examining variable documentation, frequency distributions and descriptive statistics. Particular attention was given to the difference between percentages based on all cases and valid percentages that exclude specified missing observations. The need to identify the denominator before reporting a frequency pattern was emphasized. Means, standard deviations and extreme values were discussed in relation to the level of measurement.

Data transformation included recoding categories, reversing scales and creating new variables without overwriting the original observations. Saving commands in SPSS Syntax makes transformations inspectable and repeatable rather than dependent on undocumented interface operations. Recording missing-data conventions, recoding decisions and the order of transformations was presented as a requirement for interpreting any statistical tests performed afterwards.

19 November 2025 — Descriptive statistics and statistical inference

Descriptive analysis was distinguished from inference about a population using sample observations. Mean, median and mode were compared, including their different sensitivity to skewed distributions and outliers. Range, variance and standard deviation were introduced as measures of spread that complement location measures. It was emphasized that a measure of central tendency alone could not adequately characterize dispersion.

The session developed the concepts of the normal distribution, standardization, z-scores, sampling distributions and standard errors. Standard deviation describes variation among observations, whereas a standard error describes uncertainty arising from variation across repeated samples of a statistic. The central limit theorem was discussed in relation to the distribution of sample means under suitable conditions, without assuming that the original observations were normally distributed.

Related Academy notes: Probability and Statistics

26 November 2025 — Normal approximation, the central limit theorem and reliability

Sampling distributions, the central limit theorem and the effect of sample size on estimation uncertainty were revisited. SPSS histograms and other diagnostic plots provided a context for interpreting distributional shape and potentially unusual observations. The assumption that raw data are normal was distinguished from an approximately normal sampling distribution for a particular statistic.

The session proceeded to variable transformations and scripted data operations before addressing validity and reliability in measurement. Internal consistency statistics provide information about agreement among items, but do not by themselves establish that an instrument measures the intended construct. Data preparation, distributional checks and measurement quality were considered together as prerequisites for credible statistical inference.

Related Academy notes: Probability and Statistics

3 December 2025 — Hypothesis tests, group comparisons and correlation

Null and alternative hypotheses, test statistics and p-values were examined. The p-value was distinguished from the probability that the null hypothesis was true; it was interpreted through the probability of observing a result at least as extreme under the specified null model and its assumptions. Standard deviation, standard error and confidence intervals were compared to prevent the misleading interpretation of error bars.

Independent-sample and paired t-tests, variance-assumption checks, chi-square tests for categorical associations and Pearson correlation were discussed. Selecting a test depends on measurement scale, sampling design, dependence among observations and model assumptions, not merely on an available software command. Correlation was distinguished from evidence of causation, and statistical significance from effect size.

Related Academy notes: Probability and Statistics

12 December 2025 — Python programming for data analysis

Python development environments, Jupyter workflows, package installation and module imports were introduced. Variables, numeric and textual types, lists, tuples, sets and dictionaries formed the programming foundations. Indexing and slicing, formatted output, comparisons, conditions, loops and functions were considered as building blocks of a data-processing workflow.

File input/output led into an overview of NumPy, pandas and Matplotlib. NumPy, pandas and Matplotlib were compared through their roles in numerical arrays, labeled tabular structures and plotting. While interactive notebooks make experimentation convenient, their execution order may also affect the reproducibility of results. The session connected operations already encountered in spreadsheet tools with their programmable equivalents.

19 December 2025 — Python control flow and reusable functions

The Jupyter environment and the management of isolated Python environments were revisited before examining dictionaries, sets and tuples in greater detail. Key-value access, unique-element collections and immutable tuples were discussed in terms of their different representation needs. Arithmetic and logical operations, if/elif/else branches and for/while loops provided the basis for algorithmic control flow.

Function parameters and return values were presented as a means of organizing reusable computations. Explicitly converting user input into the intended type was considered a simple but important correctness measure. These concepts established the programming prerequisites for the subsequent lessons in array processing, filtering, transformation and aggregation.

26 December 2025 — NumPy arrays and vectorized numerical computation

The session progressed from Python files and collection types to NumPy's multidimensional ndarray. Array construction, shape and dimensions, reshaping, indexing and element types were examined. Lists and homogeneous numeric arrays were contrasted in terms of memory organization and available operations. Vectorized computation illustrated how numerical work can be expressed as array operations rather than repeatedly dispatched by Python-level loops.

Aggregations including sums, means, minima, maxima and axis-based reductions were applied to multidimensional arrays. Shape compatibility and broadcasting rules were examined, including the fact that arrays of arbitrary shapes cannot necessarily be combined elementwise. NumPy was positioned as a numerical foundation for subsequent statistical workflows and machine learning.

2 January 2026 — pandas tables, indexing and missing data

The course introduced pandas Series and DataFrame structures after NumPy. Labeled rows and columns, indexing, selection and conditional filtering were used to demonstrate programmatic table processing. Missing-value detection and handling were examined with attention to type-specific representations and the risks of adopting an automatic imputation rule without considering the analysis.

Derived columns, summary statistics and basic Matplotlib plots connected data transformation to exploratory visualization. Separating selection, transformation and visualization helps reuse the same source records across different questions. Keeping source data intact and recording the steps used to produce a derived dataset were emphasized as safeguards for traceable analysis.

9 January 2026 — pandas joins, reshaping and grouped analysis

pandas Series, DataFrame and Index were reviewed through multilevel indexing (MultiIndex), stack/unstack operations and table reshaping. Axis-based concatenation using concat was distinguished from relational joining using merge. The uniqueness of join keys, unmatched records and possible many-to-many matches were examined because joins can change the number of resulting rows.

Grouping and aggregation with groupby, missing-data cleaning, descriptive summaries and file input/output were addressed. Grouped summaries and individual observations were compared as representations of information at different levels of aggregation.

Related Academy notes: Database Management Systems

Specialization — Artificial Intelligence Modules

After the common stage, I selected Artificial Intelligence and Machine Learning and Artificial Intelligence and Facilitating Tools. Their combined 23 sessions took place between 6 February and 12 May 2026. They are listed by date, with the corresponding module identified for every session.

6 February 2026 — Scope of machine learning and neural networks

Course/module: Artificial Intelligence and Machine Learning

The machine learning sequence began by situating artificial intelligence, machine learning and deep learning within their respective scopes. Rule-based software was compared with systems that estimate parameters from data. Inputs, outputs and learned mappings provided the basis for discussing generalization. Neural-network units, weights, biases, layers and nonlinear activation functions were introduced as elements of a trainable model.

Loss functions, forward propagation, gradients and backpropagation were examined in relation to parameter optimization. Performance on training examples was distinguished from performance on previously unseen observations, clarifying the difference between memorization and generalization. Biological inspiration was separated from the mathematical operation of artificial networks, while historical developments provided context for contemporary architectures.

Related Academy notes: Artificial Neural Networks

10 February 2026 — Data-processing algorithms and introduction to AI tools

Course/module: Artificial Intelligence and Facilitating Tools

The session introduced algorithms, variables and data structures as foundations for research and data-analysis tools. Arrays, records and key-value representations were related to classification, filtering and text-processing tasks. Conditionals, loops and functions were presented as ways to apply defined operations consistently across different datasets.

Input-dependent time and memory requirements were examined as factors in selecting algorithms for datasets of different sizes. Data structures and algorithms formed the programming basis for text annotation, data cleaning and operations using model APIs.

Related Academy notes: Data Structures and Algorithm Analysis

14 February 2026 — Learning paradigms and reinforcement learning

Course/module: Artificial Intelligence and Machine Learning

Supervised learning with labeled examples, unsupervised structure discovery and semi-supervised learning with limited labels were distinguished. Classification and regression were differentiated according to their targets. The relation among the problem statement, mathematical formulation and program implementation was emphasized: an appropriate model follows from a well-specified task, not from a preferred library.

Reinforcement learning introduced agents, environments, states, actions, rewards and policies. Maximizing immediate reward was contrasted with optimizing discounted future return. The exploration–exploitation tradeoff was used to explain why an agent sometimes tests less certain actions rather than repeatedly selecting its current best option. The session characterized learning methods by the feedback available to the learner as well as by model structure.

Related Academy notes: Statistical Learning and Machine Learning

20 February 2026 — Data preparation, clustering and transformer architectures

Course/module: Artificial Intelligence and Machine Learning

Web data acquisition was examined through requests, responses, parsing and structured storage. Exploratory data analysis addressed distributions, missing values and outliers. Data augmentation was discussed as a means of increasing useful variation, not as a way of creating genuinely independent observations. Supervised and unsupervised approaches were compared before examining K-means, including assignment to centroids, centroid updates and sensitivity to initialization and cluster count.

Transformer architectures and attention were introduced through context-dependent weighting across elements of a sequence. Autoregressive text generation was considered through the GPT model family, with attention to representation and context. Multimodal models were discussed as systems that may combine text with images or audio for a common task. The session connected classical analysis and clustering with newer representation-learning methods.

Related Academy notes: Statistical Learning and Machine Learning, Large Language Models

27 February 2026 — Modeling strategies, task types and overfitting

Course/module: Artificial Intelligence and Machine Learning

Automation, machine learning and deep learning were differentiated by their capabilities and scope. The main components of a learning workflow—data, target variables, features, model parameters, a loss function, an optimization method and evaluation data—were reviewed. Classification, regression and clustering were separated by output type and the availability of labels. Assumptions and computational costs associated with different model families were discussed.

Underfitting and overfitting were examined by comparing training and validation behavior. Greater model complexity was not treated as a guarantee of improved generalization. Training, validation and test partitions serve different purposes; using them interchangeably undermines evaluation. The appropriateness of an objective and its metrics was positioned as a design decision that should precede model selection.

Related Academy notes: Statistical Learning and Machine Learning

3 March 2026 — Language models, text annotation and APIs

Course/module: Artificial Intelligence and Facilitating Tools

Manual annotation and language-model-assisted categorization were compared for qualitative and textual datasets. Codebooks, category definitions, context and the consistency of examples were treated as necessary for evaluating classification results. Summarization, information extraction and label suggestions were presented as possible applications of large language models, alongside their risks of factual and contextual error.

Prompt design, context length and generation parameters were considered as factors influencing outputs. API access introduced questions of credential protection, authorization and the transfer of confidential information to external services. Fixing evaluation prompts and using a consistent reference dataset were discussed as prerequisites for reproducible comparisons in research workflows.

Related Academy notes: Large Language Models, Prompt Engineering

6 March 2026 — Classification, K-means and regularization

Course/module: Artificial Intelligence and Machine Learning

Decision boundaries, linear separability and the effects of feature scaling were examined in supervised classification. K-means was revisited through distance-based assignment and centroid updates to clarify how unsupervised clustering differs from labeled prediction. Cluster memberships should not be treated as known class labels because the objectives are distinct.

L1 and L2 regularization were discussed as mechanisms for constraining model complexity and potentially improving generalization. Scaling, fitting preprocessing operations only on training data, cross-validation and suitable metrics were reviewed. The evaluation of a classifier was considered beyond its training accuracy, including class balance, error types and performance on held-out observations.

Related Academy notes: Statistical Learning and Machine Learning

10 March 2026 — Language-model-assisted coding and classification metrics

Course/module: Artificial Intelligence and Facilitating Tools

The methodological use of language models for textual coding was examined as a sequence of defining research questions, specifying label categories, preparing annotated examples and comparing predictions against human-reviewed references. The fluent appearance of model output was distinguished from measurement validity and repeatable classification performance.

Confusion matrices introduced true positives, false positives, true negatives and false negatives. Precision, recall, F1 and accuracy were compared, particularly where classes are imbalanced or errors carry different costs. Per-class and macro-averaged results were considered together rather than reducing model quality to a single headline score. The session connected qualitative research tasks with quantitative validation.

Related Academy notes: Large Language Models, Prompt Engineering

13 March 2026 — Dimensionality reduction: PCA, LDA and SVM

Course/module: Artificial Intelligence and Machine Learning

Redundancy, correlation and discriminative information were examined in high-dimensional feature spaces. Principal component analysis (PCA) was described as an unsupervised linear projection toward directions of high variance in centered data. Linear discriminant analysis (LDA) was compared as a supervised method using class labels to seek discriminative projections. The two methods optimize different objectives and should not be treated as interchangeable.

Support vector machines (SVMs) introduced maximum-margin classification, the role of support vectors and kernel-based nonlinear decision boundaries. Scaling and projection parameters should be learned on training partitions and applied unchanged to held-out observations. This separation was emphasized as a means of preventing information leakage into model evaluation.

Related Academy notes: Statistical Learning and Machine Learning

17 March 2026 — Spatial and temporal data visualization

Course/module: Artificial Intelligence and Facilitating Tools

Visualization was considered an analytical method for examining distributions, changes and patterns rather than mere presentation. StoryMapJS supported location-based narratives, TimelineJS chronological sequences, and JuxtaposeJS side-by-side image comparisons. Historical map georeferencing introduced control points, coordinate alignment and positional uncertainty.

Interactive representations produced using tools such as Flourish were evaluated according to their chart types, scaling and data provenance. A visually persuasive map or timeline does not replace the need for a trustworthy dataset. Stating sources, time periods and transformations was presented as essential for readers who wish to interpret or verify the visual evidence.

24 March 2026 — Reproducible visualization in R and Python

Course/module: Artificial Intelligence and Facilitating Tools

Template-based tools and code-driven graphics were compared in terms of reproducibility, version control, customization and batch production. Chart forms for categorical and continuous variables were considered alongside comparisons, distributions and relationships. Chart selection was related to measurement scales, observed distributions and the research question.

R's ggplot2 grammar of layered graphics was examined alongside common Python plotting workflows. Axes, labels, scales, legends and output settings were reviewed. Diverging stacked bar charts for Likert responses require consistent ordering, treatment of neutral answers and percentages. Separating data cleaning from plotting makes results reproducible and the underlying calculations easier to audit.

Related Academy notes: Probability and Statistics

27 March 2026 — Naive Bayes, decision trees and ensemble methods

Course/module: Artificial Intelligence and Machine Learning

The Naive Bayes classifier was introduced through conditional probabilities and its assumption of conditional independence among features given a class. Decision trees were examined as recursive partitions based on feature thresholds, and random forests as ensembles that combine results from multiple trees. These approaches were compared in interpretability, variance and computational behavior.

Classification evaluation revisited confusion matrices, precision, recall, F1 and accuracy. The possibility of high overall accuracy concealing weak performance on minority classes was emphasized. Missing values, categorical coding, scaling and feature selection were discussed as preprocessing decisions that must respect the separation between training and evaluation data.

Related Academy notes: Statistical Learning and Machine Learning

31 March 2026 — AI-assisted transcription and qualitative coding

Course/module: Artificial Intelligence and Facilitating Tools

The workflow for turning recorded speech into text, checking a transcript and conducting thematic coding was examined. Descriptive codes grounded in observed text were distinguished from alternative themes and interpretive codes that require a stronger analytical judgment. Transcription errors can alter the frequency and meaning of themes, particularly when speech is interpreted without sufficient context.

Computer-assisted qualitative analysis tools and AI-generated coding suggestions were discussed with the researcher retaining responsibility for final interpretation. Transferring personal or sensitive recordings to external services was considered in terms of authorization, privacy and data protection. Versioned codebooks, human review of representative samples and clear methodological disclosure were emphasized as supports for trustworthy qualitative analysis.

Related Academy notes: Automatic Speech Recognition, Natural Language Processing

3 April 2026 — Reinforcement learning and support vector machines

Course/module: Artificial Intelligence and Machine Learning

Reinforcement learning was developed through states, observations, actions, rewards, policies and discounted return. Markov decision processes (MDPs), value and action-value functions, and Bellman relationships were introduced to express the role of future rewards. Q-learning updates and epsilon-greedy action selection illustrated how an agent balances exploration with exploiting current estimates. Fully observable state information was distinguished from partial observation.

The second part returned to SVMs. Hard and soft margins, penalties for classification violations and kernel methods for nonlinear boundaries were compared. Sequential decision-making and supervised classification differ in feedback, objectives and available observations.

Related Academy notes: Statistical Learning and Machine Learning

7 April 2026 — AI ethics, privacy and research security

Course/module: Artificial Intelligence and Facilitating Tools

AI systems were treated as sociotechnical systems shaped by data selection, model design, deployment context and human oversight. Bias, automation overreliance, unwarranted confidence and the acceptance of unverified generated claims were examined. Combining datasets can re-identify individuals even if obvious identifiers have been removed, demonstrating that privacy cannot be reduced to deleting a small set of fields.

Research practices included data minimization, authorization, confidentiality, source verification and transparent disclosure of AI assistance. Sending information to external model providers requires attention to institutional policies and applicable legal obligations. Energy and resource consumption were also discussed from a system lifecycle perspective. Technical utility was therefore assessed alongside accountability, security and environmental cost.

10 April 2026 — Deep learning, convolutional networks and attention

Course/module: Artificial Intelligence and Machine Learning

Layered representation learning, perceptrons, nonlinear activations and forward/backward propagation were reviewed. Gradients computed from the loss function were connected to optimization, learning-rate choice, convergence and overfitting. Model architectures, hardware acceleration and the amount of available training data were considered jointly when discussing practical scaling.

Convolutional neural networks (CNNs) were characterized by local operations with shared parameters. Transformer attention was contrasted as a mechanism for modeling contextual relationships among sequence elements. Both model families can be applied to multiple data types, yet they embody different inductive biases and computational costs. Learning-rate schedules and evaluation were treated as choices intended to preserve generalization, not simply decrease training loss.

Related Academy notes: Artificial Neural Networks

17 April 2026 — Generative models, tokenization and language-model training

Course/module: Artificial Intelligence and Machine Learning

Autoregressive generation, diffusion models and generative adversarial networks (GANs) were compared as distinct modeling approaches. Encoder and decoder roles, tokenization, positional information and contextual prediction were examined within transformer architectures. Statistical regularities learned from training data were distinguished from verified factual knowledge.

Pretraining, task adaptation and prompt-based conditioning were compared. Model scale, capacity and generalization behavior were discussed in relation to training and validation data. In-context learning was explained as the effect of examples supplied in a prompt, without requiring an update of the underlying model parameters. PyTorch concepts were connected to tensors and the structure of a training loop.

Related Academy notes: Large Language Models, Natural Language Processing

21 April 2026 — Text-oriented tools for data cleaning and reporting

Course/module: Artificial Intelligence and Facilitating Tools

The session examined how academic publication records can be structured, cleaned in spreadsheets and summarized through text-based instructions. A workflow involving bibliographic retrieval, Google Sheets processing and Google AI Studio-supported visualization was considered. AI-assisted spreadsheet functions and prompt-driven application generation were evaluated as techniques requiring independent checks of data schemas and outputs.

Duplicate records, missing fields and bibliographic identifiers were addressed as data-quality issues. Sharing permissions for published reports and the terms governing the underlying datasets were discussed. Automatically generated figures should not introduce findings absent from the source data, and the instructions used to create a report should remain traceable.

Related Academy notes: Prompt Engineering

28 April 2026 — AI in literature review and academic writing

Course/module: Artificial Intelligence and Facilitating Tools

Research question formulation, keyword design, literature discovery and mapping relations among publications were explored. Elicit, Connected Papers, ResearchRabbit, Google Scholar and related services were compared according to their retrieval and synthesis functions. A language-model summary was not treated as a substitute for reviewing the primary paper's methods, evidence and conclusions.

Document-grounded question answering, research-note organization and drafting tools were considered alongside the need to verify bibliographic metadata, DOIs, quotations and references. Search scope, exclusion criteria and publication date ranges must be explicit for a review to be reproducible. AI assistance was framed as support for research rather than a transfer of responsibility for interpretation or attribution.

Related Academy notes: Prompt Engineering

29 April 2026 — LLM training, quantization and deployment

Course/module: Artificial Intelligence and Machine Learning

Self-supervised pretraining, supervised adaptation, distillation and parameter-efficient fine-tuning were compared. Numerical precision was considered in relation to model memory requirements, accelerator hardware and execution costs. Multi-GPU data and model parallelism were introduced, including communication overhead in distributed computation.

LoRA/PEFT techniques, serving models and running local inference systems were discussed. Quantization can change memory use, supported operations and output quality; it is not solely a file-size optimization. Generation settings and text-classification results were evaluated in relation to fixed test sets and explicit metrics. Training, adapting and serving a model were identified as related but distinct engineering tasks.

Related Academy notes: Large Language Models

5 May 2026 — AI-supported presentation and design tools

Course/module: Artificial Intelligence and Facilitating Tools

Canva, Gamma and NotebookLM were examined for drafting slides, arranging visual material and producing source-grounded content. Automated layout and text generation were distinguished from evidence verification. Information density, visual hierarchy, readability, accurate chart interpretation and reference attribution were treated as central presentation requirements.

Generated statements were to be compared with their underlying documents, and reused charts were to retain their provenance. Visual polish was distinguished from the technical and academic reliability of a presentation. Human review, validation of visuals and institutional confidentiality constraints were therefore considered together with the convenience of automated design.

8 May 2026 — RAG, tool-using agents and local models

Course/module: Artificial Intelligence and Machine Learning

AI agents were considered through cycles of observation, decision and tool invocation. Systems that invoke external functions or resources were distinguished from text-only generation; their permissions, failure handling and termination conditions require explicit engineering. Retrieval-augmented generation (RAG) was introduced as retrieving relevant documents and placing their content in a model's context. Retrieval quality and answer quality were treated as separate evaluation dimensions.

Distillation, LoRA adaptation, local deployment using GGUF model files and visual agent-workflow tools were reviewed. Retrieval-augmented generation was distinguished from parameter training, while inaccurate, outdated or unauthorized retrieval was identified as a risk to response quality. Secure tool calls, controlled access and comparable evaluation datasets were highlighted as prerequisites for production agent systems.

Related Academy notes: Retrieval-Augmented Generation (RAG), FAISS Index Architecture

12 May 2026 — Productivity, workflow automation and office tools

Course/module: Artificial Intelligence and Facilitating Tools

Meeting transcription and summarization, calendar scheduling, knowledge organization and application-to-application automation were examined. Fathom-style meeting assistants, Reclaim scheduling tools, Notion/Notion AI and Zapier workflows illustrated different forms of process support. Automated meeting notes require verification of speakers, dates, decisions and assigned actions; a generated summary should not be accepted uncritically as an authoritative record.

Trigger-action automation was discussed in relation to connection permissions, duplicate executions, data-transfer boundaries and dependence on external services. Third-party processing of personal or institutional information requires attention to confidentiality, retention and access controls.

Digital University Platform — 2026–2027

Following my 2025–2026 Data Analysis School training, I was accepted into the YÖK Digital University Platform 2026–2027: Artificial Intelligence Training and began the next stage of my studies. I selected Artificial Intelligence: Machine Learning and Modern Software Development within the School of Artificial Intelligence.

The common classes follow the programme opening on 3 November 2026. Modern Software Development runs from 3 March to 2 June 2027, and Machine Learning from 5 March to 4 June 2027. I will record my session notes on the new page and expand the relevant Academy courses as the training progresses.

The official Digital University Platform schedule provides the current module dates and details.

Subjects and Methods Covered

The common classes provided instruction in variable types, sampling, data cleaning, descriptive and inferential statistics, hypothesis tests and programmatic data analysis. Work began with Excel and SPSS and continued with Python, NumPy and pandas.

The Machine Learning module covered model selection, classification, regression, clustering, dimensionality reduction, neural networks and large language models. The Facilitating Tools module covered text analysis, transcription, research-data preparation, visualization, literature discovery and automation. Train-test separation, appropriate metrics, source verification and human review of generated output were recurring methodological subjects across the two modules.