EGI: A Multimodal Emotional AI Framework for Enhancing Scrum Master Real-time Self-Awareness Peter Bloodsworth
Jingni Huang†
Department of Computer Science University of Oxford Oxford, United Kingdom [email protected]
Department of Computer Science University of Oxford Oxford, United Kingdom [email protected], [email protected]
While increasing research focuses on the emotional well-being of agile team members, a significant gap remains in emotion monitoring studies for Scrum Masters and meeting organizers, whose impact on team dynamics is crucial. This paper proposes a novel application integrating four carefully selected and recommended AI models to monitor the unconsciously expressed emotions of these key roles. This is achieved through: real- time transcription using a speech-to-text model; thresholding for intonation analysis to detect emotional cues in prosody; applying emotion-based vocabulary matching to identify sentiment in spoken content; and providing context-aware suggestions containing emotion keywords using an open-source, multi-module AI API. The system achieved an ASR word error rate WER of 10% in simulated meeting environments. Our evaluation shows that realtime feedback significantly improves emotion awareness during simulated agile meetings, providing Scrum Masters and meeting organizers with real-time and practical suggestions to help them quickly identify and minimize the expression of negative emotions, fostering more positive and effective team interactions.
iterative and fast-paced nature of agile often leads to team stress and emotional friction. While existing research extensively examines the collective emotional response of agile teams [47], a significant research gap remains: the emotional monitoring of the Scrum Master. As the enforcer of Scrum values, a Scrum Master’s emotional stability is critical to sustaining team morale and productivity [7]. An unaddressed display of anxiety or frustration by a leader can lead to a loss of confidence among engineers and decreased productivity. Current AI solutions often struggle with the nuanced, real-world emotional states found in professional settings. While large-scale language models excel in text processing, they frequently lack the multi-modal integration required to provide timely, actionable feedback during live interactions. Our research addresses this by introducing the EGI. Unlike existing conversational tools, EGI employs a multimodal approach that fuses automatic speech recognition (ASR [8]) with real-time prosodic analysis. By combining "what" is said (linguistic content) with "how" it is said (intonation), EGI provides Scrum Masters with immediate visual and textual feedback, facilitating self-awareness and fostering a positive team environment without the intrusive need for videobased monitoring.
CCS CONCEPTS
2
ABSTRACT
• Software and its engineering → Software development process management • Computing methodologies → Machine learning • Human-centered computing → Interactive systems and tools
KEYWORDS Multimodal AI, Agile Software Development, Scrum Master, Affective Computing, Speech Recognition, Sentiment Analysis ACM Reference format: Jingni Huang and Peter Bloodsworth. 2026. EGI: A Multimodal Emotional AI Framework for Enhancing Scrum Master Real-time Self-Awareness. In Proceedings of the 1st International Workshop on AI-powered Software (AIware ’26). https://doi.org/10.1145/nnnnnnn.nnnnnnn, Montreal, Canada
1 Introduction Agile software development [1] emphasizes effective communication at the heart of project success[2-6]. However, the
Methodology
Our approach integrates Human-Computer Interaction (HCI) design principles with rigorous signal processing to meet the lowlatency requirements of real-time agile ceremonies. To ensure high-performance speech-to-text (STT) integration, we conducted an extensive evaluation of deep learning architectures, including RNN, LSTM, and Transformer-based models. We specifically compared Wav2Vec 2.0 and OpenAI’s Whisper model. Our experimental validation using established datasets, such as the Speech Emotion Recognition Voice Dataset [44] and the Opinions Lexicon [45], demonstrated that while Wav2Vec offers low latency, Whisper provides superior Word Error Rate (WER) performance in noisy, multi-accent meeting environments. Consequently, Whisper was selected for the linguistic pipeline.
AIware '26, May 15, 2026 Montreal, Canada
3
Application System Architect
J.Huang and P. Bloodsworth
advanced users. The ultimate goal of the entire system is to help Scrum Masters quickly analyze their sentiment (positive or negative) and provide immediate response recommendations without increasing the burden on the Scrum Master.
3.3
Signal Processing and Data Flow
Figure 1: System Architecture(User Video track is future work)
From figure 1 system architecture we could see the application is divided into 4 functional models. First model speech-to- text: Speech-to-Text: This is the first step, using ASR to get the transcript text to convert spoken words into text, laying the foundation for subsequent analysis.
3.1
Model Evaluation and Selection
Second mode intonation visualization: Decode the original audio to plot pitch and Amplitude for visualization as the figure 1 Demo’s first and second dynamics visualization. Visualization can immediately grab the eye's attention and form a sharp contrast with the continuous waveform of voice intonation, with sudden high and low pitches. Third model emotion keyword prompt: Transcript text from first model speech to text as input for matching the sentiment word library of positive and negative sentiment words, then display the result on the keyword display bar. It only shows 3- 5 words which reduce reading time and improve first glance recognition under the intonation visualization. Fourth model suggestions: This is an optional enhancement module. Using the first model text output for combining with multiModal model gpt-4o-transcribe module open source in the end for more complex sentiment analysis and response suggestions, if the scrum master requests extra support. The system architecture separates core real-time feedback functionality (tone visualization and short keywords) from complex on-demand analysis (recommendations). Because the application requires only microphone input, not a camera, this design ensures real-time performance while balancing cost- effectiveness and user comfort. Advanced multimodal options are also available for
Figure 2: Application Data Flow Diagram (User Video track is future work)
The EGI framework optimizes the inference pipeline through a decoupled, resource- aware task allocation strategy. As illustrated in the data flow (Figure 2), the system processes multimodal inputs via two primary parallel paths to capture a holistic emotional profile: Path A (Linguistic Path): This path utilizes the Whisper model to generate high-fidelity transcripts from raw audio. These transcripts are then passed to the GPT-4o API for advanced semantic inference and the generation of context-aware suggestions. This ensures that the system leverages the superior reasoning capabilities of Large Multimodal Models (LMMs) for complex emotional support. Path B (Acoustic Path): Simultaneously, this path performs local Digital Signal Processing (DSP) using libraries such as librosa or Aubio. It extracts key prosodic features, specifically the fundamental frequency (F0) for pitch and root mean square (RMS) energy for amplitude. By processing these features locally, the system ensures the delivery of time-sensitive emotional cues (intonation) with zero network latency**, providing a responsive and fluid user experience. This dual-path strategy enables EGI to capture emotional valence that pure text analysis typically misses-such as the discrepancy between a positive transcript and a negative tone. Furthermore, the EGI framework is designed for scalability. While the current implementation prioritizes acoustic and linguistic modalities to maintain low latency and user privacy, the architecture reserves a User Video track (as shown in Figure 1) for future integration of facial expression analysis, which will further enrich the multimodal emotional profile.
EGI: A Multimodal Emotional AI Framework for Enhancing Scrum Master Real-time Self-Awareness
4 4.1
Experimental Evaluation Model Selection and Benchmarking Table 3: Wav2Vec 2.0 with Whisper Comparison[30]
AIware '26, May 15, 2026 Montreal, Canada contains audio clips categorized by four distinct emotions, providing the foundational data for our intonation and ASR experiments. The latter is the core of our emotion tagging model. The performance of the speech-to-text module is evaluated using the Word Error Rate (WER), defined as: WER = (S + D + I) / N (1) Where S is the number of substitutions, D is deletions, I is insertions, N is the total number of words in the reference transcript. Keep leanings and potential optimization include: 1. Implementing noise reduction techniques to achieve strong realtime performance; 2. Practice prompt engineering to optimize the output of the AI suggestion model. Table 3: models performance and comparison Model
According to literature benchmarks [30][44][45], Whisper demonstrates superior robustness compared to Wav2Vec 2.0 in handling diverse accents and noisy environments typical of agile meetings. While Wav2Vec 2.0 offers slightly faster inference as denoted as table 3, we selected Whisper to prioritize the textual accuracy required for reliable downstream sentiment analysis.
4.2
Speech-to-text Whisper transcription
Accurac y/Result
Key Challen ges
Proposed Optimizations
WER: 10%
Small dataset with varied intonati ons and noisy; high WER
pre-process noise, refine kenLM integration
Subject ive interpre tation of pitch levels
Threshold as reference
Integrate Intonation visualization model to text dataset; NLP models (e.g., BERT) for nuanced sentiment
Cache frequent queries; fine-tune local LLM for emotion-specific suggestions
Databases
Speech Emotion Recognition Dataset [44]: We chose this concise dataset because although it contains only 68 audio samples, it contains accurately labeled and transcribed text, and the audio contains four different emotional expressions, which fully meets the database requirements for our initial experiments. These texts are read aloud in English representing four different emotional states: euphoria, joy, sadness, and surprise. Each audio clip captures the speaker‘s tone, intonation, and nuances in expressing emotions through sound. This dataset was chosen because the audio clips themselves capture the different intonations, nuances, and subtleties of emotional speech, making them ideal for validating the robustness of our system in unpredictable and emotional environments such as agile meetings. The name of each audio clip itself provides a keyword for true emotion. Opinion Lexicon[45]: A sentiment lexicon containing 2006 positive and 4783 negative English words, used in keyword suggestion models to identify sentiment words in transcribed text.
4.3
ASR
Approach
Experimental Results and Discussion
Our end-to-end evaluation aimed to demonstrate the effectiveness of an ensemble of four models using a pre-prepared dataset. To this end, we leveraged two key datasets: the Speech Emotion Recognition Dataset and the Opinion Dictionary. The former
Intona tion visuali zation
keyw ord prom pt
AI Sugg estion Gener ator
Waveform and pitch visualization
Qualitati ve pitchemotion correlati on
Text matching with opinion lexicon
String prompts
Limited to lexical sentim ent; misses context ual tone
GPT-4o for further feedback
Context -aware respons es
Depen dency on externa l APIs; latency
As table 3 models performance and comparison shows the models experiment comparison: 1. Speech-to-Text (ASR): Whisper This experiment uses the Whisper speech- to-text model and decodes it to accelerate processing.Preliminary results show a word
AIware '26, May 15, 2026 Montreal, Canada
error rate (WER) of 10%. WER is a key metric for measuring the performance of speech recognition systems. A 10% WER means that 10 out of every 100 words are misrecognized, inserted, or deleted. This result indicates that while the model works, its accuracy still has significant room for improvement. Major Challenges: The primary challenge stems from the dataset’s limitations. While the dataset was chosen to ensure model robustness, its small size and wide range of intonations and background noise significantly interfere with the model's recognition accuracy. The small dataset limits the model's learning and generalization capabilities, while the varying intonations and noise increase the difficulty of recognition. The high WER is a direct reflection of these challenges. Optimization Solutions: To address these challenges, the experiment proposed two main optimization measures. Preprocessing Noise: By reducing the audio noise, we can provide the model with cleaner input, thereby improving its recognition accuracy. Optimizing KenLM Integration: KenLM is an efficient language model toolkit used to improve decoding results in speech recognition. 2. Intonation Visualization: Waveform and Pitch Visualization This section aims to visualize the waveform and pitch of speech to qualitatively analyze the relationship between pitch and emotion. By converting audio data into waveform graphs and pitch curves, researchers can visually observe pitch variations and subjectively infer their correlation with the speaker's emotion. For example, a sharp rise in pitch may indicate surprise or anger. Key Challenge: A core challenge of this approach lies in the subjective nature of the pitch-emotion association. While pitch can provide emotional clues, its interpretation is highly dependent on the speaker's cultural background, context, and individual perception. Objective and accurate emotional judgment based solely on pitch level is difficult, as the same emotion (such as excitement) can manifest differently in different speakers or contexts. Optimization Solution: To address this subjectivity, the project proposes a simple optimization method: setting a threshold as a reference. For example, a baseline pitch value can be set, and any pitch variation exceeding this threshold is considered emotionally expressive, providing a relatively objective starting point for qualitative analysis. 3. Sentiment keywording: Text Matching and Sentiment Dictionaries To assign sentiment keywords to text content, the project employed a text matching method based on sentiment dictionaries. This method matches the input transcription against a predefined list of sentiment terms. For example, text containing words like "good" and "like" might be keyworded as positive. This method can complement the tone visualization model described above. When
J.Huang and P. Bloodsworth
changes in speech tone are insufficient to convey emotion, the context of the conversation can be further referenced. Main Challenge: This method has significant limitations. It is limited to sentiment analysis at the lexical level and ignores context. Many words are inherently neutral but can take on strong emotional overtones in specific contexts. Furthermore, complex linguistic phenomena such as irony and puns cannot be identified through simple lexicon matching, resulting in the analysis results failing to capture the true emotional tone of the text. However, future research will focus on this aspect and further data testing in conjunction with the tone visualization model described above. Optimization: To improve the accuracy and depth of sentiment analysis, the project recommends integrating context-aware natural language processing (NLP) models, such as BERT. And as mentioned earlier, it is combined with the previous intonation visualization to perform data testing and parameter optimization. 4. AI Recommendation Generation: Using GPT-4o This project leverages the powerful generative capabilities of GPT4o to provide contextual recommendations. GPT-4o is a state-ofthe-art large language model (LLM) that can understand and generate high- quality, contextually relevant text. By taking previously transcribed text and camera- projected video as input, GPT-4o can provide more insightful and targeted feedback or recommendations. Key Challenges: The primary challenge with this approach lies in its reliance on external APIs and the resulting latency. Each call to GPT-4o requires a request to an external server, which not only incurs costs but can also slow down application response times due to network latency and server response times. For applications requiring real-time feedback, this latency needs to be further improved. Optimization: Caching Frequent Queries: For common or repeated queries, the results can be cached. When the same query is encountered again, the answer can be retrieved directly from the local cache without making another call to the external API. This can significantly reduce latency and API call costs. Fine-tuning a local LLM for specific emotions: This project proposes using a smaller, more lightweight local large language model and fine-tuning it for specific emotion-related recommendations. While the local LLM may not be as general as GPT-4o, through specialized training, it can perform well on specific tasks while avoiding all the problems of external APIs and achieving fast and low-cost inference. The ultimate goal of these experiments is to verify that all components meet their established performance targets - such as achieving a relative low WER and maintaining real-time processing latency - thereby confirming the overall system's practical efficacy.
5
Demo Interface
EGI: A Multimodal Emotional AI Framework for Enhancing Scrum Master Real-time Self-Awareness
Figure 4: Pitch Tracking
Figure 4 pitch tracking displays the fundamental frequency fluctuations, which dreamily correlate with emotional state(e.g. high pitches with joy or anxiety). This feature extraction confirms the functional requirement of real-time intonation visualization. The visualization aims to provide a real-time representation of the pitch changes, associating high pitches with active emotions (joyfull), mid-range with neutral emotions, and low pitches with awake emotions as referenced as [45]. As concluded in [45] prosody attributes high pitch levels to emotions, such as joy, anxiety, or fear, because high pitch draws the listener’s attention by establishing a contrast. Medium pitch levels account for more neutral attitudes. However, low pitch levels are related to sober emotions: sadness, calmness, or security. Thus, emotions carrying a high level of activity, such as joy or fear, tend to be situated at the top end of the frequency spectrum of the speaker. This experiment consists of two stages: First, the audio archive is transcribed into text using OpenAI's Whisper model, and its output is directly used as input for the second stage. Subsequently, under the guidance of specific system prompts, the transcribed text is input into the GPT-4o model to summarize the core emotions and analysis. The final text output then contains this summary and a suggested solution. It should be noted that there are costs associated with using GPT-4o.
AIware '26, May 15, 2026 Montreal, Canada Although the text content was the same, the sentence was spoken with two distinct tones. This difference is clearly visualized in the audio amplitude and pitch graphs, despite the consistent keyword tags. The final output confirms this, concluding with "Joyful" for the first sample and "Sad" for the second. This demonstration video highlights the application's real-time context and low- latency performance. The final conclusion, along with the visualization updates, is delivered just as, or a few seconds after, the sentence ends.
6
Conclusion
This paper presents EGI, a novel, lightweight multimodal AI application designed to enhance human emotion verification in agile software development. By integrating real-time ASR, intonation visualization, and lexicon-based keyword prompting, the system provides Scrum Masters with critical emotional insights during team ceremonies. Our experimental evaluation confirms the technical viability of the ensemble architecture, with the Whisperbased ASR module achieving a 10% WER on emotional speech datasets. The research demonstrates that a modular, microservices-based approach allows for scalable emotion tracking—balancing free, local real-time feedback with on-demand advanced AI recommendations via GPT-4o. While the current iteration successfully bridges a critical gap in Scrum communication, future work will focus on fine-tuning the ASR models with domain specific datasets, implementing advanced noise reduction, and exploring local LLM integration to reduce latency in AI-driven suggestions. Ultimately, this work provides a foundational framework for applying multimodal AI to foster team selfawareness and communication efficiency in industry standard agile environments.
ACKNOWLEDGMENTS The author would like to express their sincere gratitude to Dr. Peter Bloodsworth for his invaluable guidance, academic supervision, and continuous support throughout the development of this research at the University of Oxford. The author also thanks the agile practitioners whose feedback informed the system requirements and the broader academic community for the opensource datasets and models that made this work possible.
Figure 5: Demo Video
In addition to completing the experiments and demos above, we created a demonstration video Figure 5 using two randomly selected audio samples with the same content: "I enjoy taking leisurely strolls through the quiet countryside." This video showcases the visual visualization and the effects of different audio and intonation.
REFERENCES [1] Al-Saqqa, S., Sawalha, S., & AbdelNabi, H. (2020). Agile software development: Methodologies and trends. International Journal of Interactive Mobile Technologies, 14(11). [2] Luong, T.T., Sivarajah, U., & Weerakkody, V. (2021). Do Agile Managed Information Systems Projects Fail Due to a Lack of Emotional Intelligence? Information Systems Frontiers, 23(2), 415–433.
AIware '26, May 15, 2026 Montreal, Canada
[3] Ahmed, A., Ahmad, S., Ehsan, N., Mirza, E., & Sarwar, S. Z. (2010). Agile software development: Impact on productivity and quality. In 2010 IEEE International Conference on Management of Innovation & Technology (pp. 287-291). IEEE. [4] Edison, H., Wang, X., & Conboy, K. (2022). Comparing Methods for Large-Scale Agile Software Development: A Systematic Literature Review. IEEE Transactions on Software Engineering, 48(8), 2709-2731. doi: 10.1109/TSE.2021.306903. [5] Lowell, K.R. (2023). Agile Principle 6: “The Most Efficient and Effective Method of Conveying Information to and Within a Development Team Is Face-to-Face Conversation”. In: Leading Modern Technology Teams in Complex Times. Future of Business and Finance. Springer, Cham. [6] Bhalerao, S., & Ingle, M. (2010). Analyzing the modes of communication in agile practices. In 2010 3rd International Conference on Computer Science and Information Technology (pp. 391-395). Chengdu, China. [7] Kristensen, S. H., & Paasivaara, M. (2021). What Added Value Does a Scrum Master Bring to the Organisation? — A Case Study at Nordea. In 2021 47th Euromicro Conference on Software Engineering and Advanced Applications (SEAA) (pp. 270-278). Palermo, Italy.
J.Huang and P. Bloodsworth
[8] Torchaudio Contributors. (2024). ASR INFERENCE WITH CUDA CTC DECODER. [9] Mehriban, A. (1968). Communication without words. Psychology Today, 2(4), 53-56. [10] Fasel, B., & Luettin, J. (2003). Automatic Facial Expression Analysis: A Survey. Pattern Recognition, 36, 259-275. [11] Pathak, S., & Arun K. (2011). Recognizing emotions from speech. In 3rd International Conference on Electronics Computer Technology (ICECT). Vol. 4. [12] Gilke, M., Kachare, P., Kothalikar, R., Rodrigues, V. P., & Pednekar, M. (2012). MFCC-based Vocal Emotion Recognition Using ANN. In International Conference on Electronics Engineering and Informatics (ICEEI). IPCSIT vol. 49, IACSIT Press. [13] Rao, K. S., Kumar, T. P., Anusha, K., Leela, B., Bhavana, I., & Gowtham, S.V.S.K. (2012). Emotion Recognition from Speech. International Journal of Computer Science and Information Technologies (IJCSIT), 3(2), 3603-3607. [14] Aouani, H., & Ben Ayed, Y. (2020). Speech emotion recognition with deep learning. Procedia Computer Science, 176, 251-260. [15] Koolagudi, S.G., & Rao, K.S.(2012) Emotion recognition from speech: a review. Int J Speech Technol 15, 99–117. [16] Rodero, E.(2022). Effectiveness, attractiveness, and emotional response to voice pitch and hand gestures in public speaking. Frontiers in communication 7: 869084. [17] Burgoon, J. K., Guerrero, L. K., and Floyd, K. (2010). Nonverbal Communication. Routledge. [18] Jankowska, A., et al(2023). Enough is enough: how much intonation is needed in the vocal delivery of audio description?. Perspectives 31(4): 705723. [19] Madampe, K., Hoda, R., & Grundy, J. (2023). A Framework for EmotionOriented Requirements Change Handling in Agile Software Engineering. IEEE Transactions on Software Engineering, 49(5), 3325-3343. [20] Shastri, Y., Hoda, R., & Amor, R. (2021). Spearheading agile: the role of the scrum master in agile projects. Empirical Software Engineering, 26(1), 3. [21] Humphrey, R. H. (2002). The many faces of emotional leadership. The Leadership Quarterly, 13(5), 493-504. [22] Lipton, Z. C., Berkowitz, J., & Elkan, C. (2015). A critical review of recurrent neural networks for sequence learning. arXiv preprint arXiv:1506.00019. [23] Al-Selwi, S. M., et al.(2024). RNN-LSTM: From applications to modeling techniques and beyond—Systematic review. Journal of King Saud UniversityComputer and Information Sciences. 102068. [24] Cho, K., et al(2014). Learning phrase representations using RNN encoder-decoder for statistical machine translation. arXiv preprint arXiv:1406.1078.
EGI: A Multimodal Emotional AI Framework for Enhancing Scrum Master Real-time Self-Awareness
AIware '26, May 15, 2026 Montreal, Canada
[25] O'shea, K., & Nash, R. (2015). An introduction to convolutional neural networks. arXiv preprint arXiv:1511.08458.
[44] KUCEV, R. (2023). Speech Emotion Recognition Voice Dataset. Kaggle.
[26] He, K., et al.(2016). Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition.
[45] Liu, B. (2022). Opinion spam detection. In Sentiment analysis and opinion mining (pp. 113-125). Cham: Springer International Publishing.
[27] Vaswani, A., et al.(2017). Attention is all you need. In Advances in neural information processing systems 30. [28] Schneider, S., et al(2019). wav2vec: Unsupervised pre-training for speech recognition. arXiv preprint arXiv:1904.05862. [29] Baevski, A., et al.(2020). wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems 33: 12449-12460 [30] Radford, A., et al(2023). Robust speech recognition via large-scale weak supervision. International conference on machine learning. PMLR. [31] Guo, D., et al.(2025). Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948. [32] Sebe, N., Cohen, I., & Huang, T. S. (2005). Multimodal emotion recognition. In [Eds.], Handbook of Pattern Recognition and Computer Vision (pp. 387-409). World Scientific. [33] OpenAI. (2025). Introducing our next-generation audio models. OpenAI Blog. Retrieved from https://openai.com/index/introducing-our-nextgeneration-audio-models/. [34] Hurst, Aaron, et al.(2024). "Gpt-4o system card." arXiv preprint arXiv:2410.21276. [35] Baltrušaitis, T., Robinson, P., & Morency, L.P. (2016). Openface: an open source facial behavior analysis toolkit. In 2016 IEEE winter conference on applications of computer vision (WACV). IEEE. [36] NVIDIA Blogs.(2023, June 14). "What Is MLOps?" NVIDIA Blog. Retrieved from blogs.nvidia.com/blog/what-is-mlops/. [37] Sculley, D., et al.(2015). Hidden technical debt in machine learning systems. Advances in neural information processing systems 28. [38] Paleyes, A., Urma, R.G., & Lawrence, N.D. (2022). Challenges in deploying machine learning: a survey of case studies. ACM computing surveys, 55(6), 1-29. [39] GoogleCloud. (2024). MLOps: A guide to the machine learning lifecycle. [40] AWS. (2024). Operationalizing Machine Learning (MLOps). [41] MLOps.Machine Learning Operations, MLOps. Retrieved from https://mlops.org. [42] Amazon Web Services, Inc. (2024). What is DevOps? Retrieved from aws.amazon.com/tw/devops/what-is-devops/ [43] Google Cloud.(2025).Deploy a containerized application to Cloud Run using Cloud Build. Retrieved from https://docs.cloud.google.com/build/docs/deploy-containerized-applicationcloud-run.
[46] Kreuzberger, D., Kühl, N., & Hirschl, S. (2023). Machine Learning Operations (MLOps): Overview, Definition, and Architecture. IEEE Access, 11, 31866-31879. doi: 10.1109/ACCESS.2023.3262138. [47] Alzoubi, Y., & Gill, A. (2021). The Critical Communication Challenges Between Geographically Distributed Agile Development Teams: Empirical Findings. IEEE Transactions on Professional Communication, 64(4), 322337. [48] Gupta, S. (2025). The Rise of Serverless AI: Transforming Machine Learning Deployment. European Journal of Computer Science and Information Technology, 13(5), 45-67.