Ongoing research

Toward a Data-Centric Trust Pipeline for Trustworthy AI. Investigates how integrity, bias, synthetic data generation and provenance form an interdependent ecosystem for building trustworthy AI.

Objective

Propose the Data-Centric Trust Pipeline as a unified model to raise transparency, interpretability and accountability across the data lifecycle.

Problem lines

• How to ensure consistency, contextualization and completeness when datasets lack metadata and standardization?
  Rationale. Integrity failures compromise reproducibility and auditability, amplifying downstream errors.
• Which approaches mitigate disparities across demographic groups when fairness metrics are incomplete?
  Rationale. Studies show 20–35% error variation across groups: fairness demands continuous treatment.
• How to validate the authenticity and usage limits of synthetic data, which may reproduce structural bias?
  Rationale. Without traceability, synthetic data distorts interpretation in sensitive analyses.
• How do provenance systems evolve from passive repositories into interpretive records of decisions?
  Rationale. Provenance is essential for explainability and auditing in AI pipelines.

References

• D. Schwabe et al., "The METRIC-framework for assessing data quality for trustworthy AI in medicine: A systematic review," npj Digit. Med., vol. 7, no. 1, art. no. 11, 2024, doi: 10.1038/s41746-024-01196-4.
• J. Buolamwini and T. Gebru, "Gender shades: Intersectional accuracy disparities in commercial gender classification," in Proc. 1st Conf. Fairness, Accountability and Transparency, in Proceedings of Machine Learning Research, vol. 81, 2018, pp. 77–91.
• M. A. S. Hameed, A. M. Qureshi, and A. Kaushik, "Bias mitigation via synthetic data generation: A review," Electronics, vol. 13, no. 19, art. no. 3909, 2024, doi: 10.3390/electronics13193909.
• M. Ahmed et al., "Data provenance in healthcare: Approaches, challenges, and future directions," Sensors, vol. 23, no. 14, art. no. 6495, 2023, doi: 10.3390/s23146495.
• S. Longpre et al., "A large-scale audit of dataset licensing and attribution in AI," Nature Mach. Intell., vol. 6, no. 8, pp. 975–987, Aug. 2024, doi: 10.1038/s42256-024-00878-8.

An Architecture-Aware Conceptual Framework for Cloud Capacity Planning. Integrates architectural characteristics, workload semantics and operational objectives (SLO, PLO, RLO) into a unified conceptual model.

Objective

Propose C2PF as an architecture-aware framework capable of producing proactive, explainable and repeatable capacity decisions.

Problem lines

• How to move past reactive models based only on historical metrics, which miss variation and concurrency?
  Rationale. The absence of architectural context leads to inaccurate forecasts and overprovisioning.
• How to combine architectural components, workload classification and performance objectives?
  Rationale. There are semantic and cognitive gaps between architects, operators and planners.
• How to quantify the effects of patterns (microservices, event-driven, serverless) on capacity consumption?
  Rationale. Architectural factors can add up to 30% overhead, altering scalability and latency.

References

• C. D. Cavalcanti Pereira, "Capacity planning of cloud computing workloads," Doctoral dissertation, CESAR School, Recife, Brazil, 2023, doi: 10.2139/ssrn.5754002.
• C. D. Cavalcanti Pereira, "Capacity planning of cloud computing workloads," in Cloud Computing — Applications and Sustainable Developments. London, U.K.: IntechOpen, 2025, doi: 10.5772/intechopen.1011100.
• M. Richards and N. Ford, Fundamentals of Software Architecture: An Engineering Approach. Sebastopol, CA, USA: O’Reilly Media, 2020.
• R. Lichtenthäler, J. Fritzsch, and G. Wirtz, "Cloud-native architectural characteristics and their impacts on software quality: A validation survey," in Proc. IEEE Int. Conf. Service-Oriented System Eng. (SOSE), 2023, arXiv:2306.12532.
• N. J. Gunther, Guerrilla Capacity Planning: A Tactical Approach to Planning for Highly Scalable Applications and Services. Berlin, Germany: Springer, 2010, doi: 10.1007/978-3-540-31010-5.

Capital Intensity, Market Dynamics, and Competitive Differentiation Across the Value Chain. Analyzes how the layers — semiconductors, cloud, models, inference, platforms and applications — organize competition, investment and value capture.

Objective

Propose a conceptual model of the generative-AI value chain and evaluate hypotheses on capital intensity, infrastructure concentration, capability diffusion and productivity impacts.

Problem lines

• Economic structure and capital intensity — how the distribution of capital across layers shapes competition.
• Infrastructure concentration — the extent to which supply shows structural concentration.
• Capability diffusion — how quickly frontier capabilities diffuse.
• Application entry barriers — how APIs and platforms influence development.
• Productivity impacts — evidence of gains in knowledge work.

References

• Y. LeCun, Y. Bengio, and G. Hinton, "Deep learning," Nature, vol. 521, no. 7553, pp. 436–444, May 2015, doi: 10.1038/nature14539.
• J. Kaplan et al., "Scaling laws for neural language models," 2020, arXiv:2001.08361.
• J. Hoffmann et al., "Training compute-optimal large language models," 2022, arXiv:2203.15556.
• E. Brynjolfsson, D. Li, and L. R. Raymond, "Generative AI at work," Nat. Bureau Econ. Res., Cambridge, MA, USA, Working Paper 31161, 2023, doi: 10.3386/w31161.

A systematic literature review. Critically analyzes optimized scheduling algorithms for Kubernetes, focusing on applicability, benefits and challenges in general computing, AI and edge.

Problem lines

• How do optimized algorithms behave in production compared with the experimental tests in the literature?
  Rationale. Most studies occur in controlled environments; there is a gap on practical applicability.
• How do multi-criteria algorithms improve allocation in heterogeneous clusters (CPU/GPU, edge/cloud)?
  Rationale. The combined effectiveness of latency, GPU and energy is still uncertain.
• Which ML models are most effective at forecasting demand and adapting the scheduler?
  Rationale. Systematic comparisons of QoS and cost-benefit are lacking.

References

• G. El Haj Ahmed, F. Gil-Castiñeira, and E. Costa-Montenegro, "KubCG: A dynamic Kubernetes scheduler for heterogeneous clusters," Softw. Pract. Exp., vol. 51, no. 2, pp. 213–234, 2021, doi: 10.1002/spe.2898.
• I. Harichane, S. A. Makhlouf, and G. Belalem, "KubeSC-RTP: Smart scheduler for Kubernetes platform on CPU-GPU heterogeneous systems," Concurrency Comput. Pract. Exp., vol. 34, no. 21, art. no. e7108, 2022, doi: 10.1002/cpe.7108.
• M. Carvalho and D. F. Macedo, "Container scheduling in co-located environments using QoE awareness," IEEE Trans. Netw. Serv. Manage., vol. 20, no. 3, pp. 3247–3260, 2023, doi: 10.1109/TNSM.2023.3244090.
• T. Menouer, "KCSS: Kubernetes container scheduling strategy," J. Supercomput., vol. 77, no. 5, pp. 4267–4293, 2021, doi: 10.1007/s11227-020-03427-3.