Nariman Mani's Technical Mind

Nariman Mani, P.Eng., PhD Computer and Software Engineering

Multi-Agent Change Impact Analysis and Test Optimization for AI-Enabled Software Systems

SENTRY: Drift-Aware Validation Placement for AI Services in the Cloud-Edge Continuum

Secure AI Orchestration for Digital Health: Private Analytics and Adaptive Group Coaching

PRISM-Coach: Privacy-by-Design Adaptive Group Assignment for Lifestyle Coaching at Scale

Key Developer Roles and Organizational Coupling in Microservices: A Longitudinal View


# Multi-Agent Change Impact Analysis and Test Optimization for AI-Enabled Software Systems

## Publication at a glance

- **Paper:** *Multi-Agent Change Impact Analysis and Test Optimization for AI-Enabled Software Systems*
- **Authors:** Nariman Mani, Amr S. Abdelfattah, Shakthi Weerasinghe, Xiaozhou Li, and Tomas Cerny
- **Status:** Accepted for publication
- **Venue:** 8th IEEE International Conference on Artificial Intelligence Testing (IEEE AITest 2026), part of the IEEE CISOSE 2026 Congress
- **Location and dates:** Fukuoka, Japan, July 27–30, 2026
- **Conference:** [IEEE AITest 2026](https://cisose.fit.ac.jp/aitest/)

## The research problem

AI-enabled systems combine application code, models, data pipelines, prompts, services, and infrastructure. A change in any one of these elements can affect behavior elsewhere, making conventional file-level impact analysis incomplete. At the same time, running every available test after every change is slow and expensive.

The central challenge is therefore twofold: identify the parts of an AI-enabled system that a change can influence, and select the tests that provide the strongest evidence about those risks.

## The paper's central contribution

This work investigates a multi-agent approach in which specialized agents collaborate on change impact analysis and test optimization. Rather than asking one component to interpret every signal, the approach separates responsibilities and combines their findings into an actionable testing decision.

That separation is useful in heterogeneous systems. Source changes, model dependencies, service relationships, and test histories provide different kinds of evidence and can be examined by agents suited to each concern.

## How the approach works

At a high level, the workflow connects three activities:

1. **Change interpretation** identifies what changed and gathers relevant technical context.
2. **Impact reasoning** traces likely effects through the dependencies of the AI-enabled system.
3. **Test optimization** prioritizes tests according to the identified impact and risk.

The resulting process focuses validation effort without treating test reduction as the only goal. The more important objective is to preserve confidence while directing limited test resources toward the most relevant behavior.

## Why this publication matters

Change impact analysis is especially difficult when system behavior depends on learned components and evolving data as well as deterministic code. A collaborative agent architecture offers a way to organize that complexity, retain traceable intermediate reasoning, and adapt test selection to the nature of each change.

The work is relevant to regression testing, continuous integration, AI engineering, and quality assurance for systems whose dependencies cross traditional software and machine-learning boundaries.

## Scope and responsible interpretation

Optimized test selection should not become an excuse to ignore unmodeled risk. Dependency information can be incomplete, agent conclusions can be uncertain, and novel failures may fall outside historical evidence. Practical use therefore requires conservative fallbacks, auditable decisions, coverage monitoring, and periodic full regression runs.

Readers should consult the published paper for the precise architecture, research method, evaluation setup, and results. Bibliographic details such as a DOI and page numbers will be added after IEEE assigns them.

## How to cite this paper

> N. Mani, A. S. Abdelfattah, S. Weerasinghe, X. Li, and T. Cerny, “Multi-Agent Change Impact Analysis and Test Optimization for AI-Enabled Software Systems,” in *Proceedings of the 8th IEEE International Conference on Artificial Intelligence Testing (AITest 2026)*, Fukuoka, Japan, 2026. Accepted for publication.

Before submitting a manuscript, verify the latest metadata on the [IEEE AITest 2026 conference site](https://cisose.fit.ac.jp/aitest/). Please cite the original publication rather than this explanatory post.

## Related research on this weblog

- Browse the complete [Nariman Mani publications list](/publications).
- Read about [dependency-aware adaptive test healing with graph neural networks](/posts/gnn-dependency-aware-adaptive-test-healing).
- Explore [adaptive test healing with LLMs and reinforcement learning](/posts/adaptive-test-healing-llm-reinforcement-learning).



An overview of multi-agent change impact analysis and risk-aware test optimization for complex AI-enabled software systems.


# SENTRY: Drift-Aware Validation Placement for AI Services in the Cloud-Edge Continuum

## Publication at a glance

- **Paper:** *SENTRY: Drift-Aware Validation Placement for AI Services in the Cloud-Edge Continuum*
- **Authors:** Nariman Mani, Xiaozhou Li, Shakthi Weerasinghe, and Tomas Cerny
- **Status:** Submitted to the Cloud-Edge Continuum Workshop (CEC 2026), held with IEEE ICNP 2026
- **Workshop date and location:** October 5, 2026, Tempe, Arizona, USA
- **Workshop:** [Cloud-Edge Continuum Workshop 2026](https://cecf.github.io/cec26/)

## The research problem

AI services deployed across cloud and edge environments face a moving target. Input distributions and operating conditions can drift, while bandwidth, latency, compute capacity, privacy constraints, and connectivity differ by location. Validation that always runs in one fixed place can respond too slowly, consume unnecessary resources, or miss important local behavior.

The challenge is to decide where validation should happen as conditions change: close to an edge workload, in a more capable cloud environment, or across both.

## The paper's central contribution

SENTRY investigates drift-aware validation placement for AI services in the cloud-edge continuum. Its core idea is that evidence of drift should inform not only *whether* a service needs validation, but also *where* that validation can be performed effectively under current system constraints.

This connects AI quality monitoring with distributed-systems resource decisions. Validation becomes an adaptive operational capability rather than a fixed stage that occurs only before deployment.

## How the approach works

The high-level workflow links three concerns:

1. **Drift awareness** observes changes in the data or behavior relevant to an AI service.
2. **Placement reasoning** considers cloud-edge constraints such as latency, resources, and communication cost.
3. **Validation execution** places the appropriate checks where they can provide timely, useful evidence.

By reasoning across these concerns, a system can react to changing conditions without assuming that either the cloud or edge is always the best validation location.

## Why this work matters

Cloud-edge AI systems must balance responsiveness with assurance. Local validation may reduce latency and preserve data locality, while cloud validation may provide stronger compute capacity and broader context. Drift makes that tradeoff dynamic.

The work is relevant to continuous validation, MLOps, edge intelligence, adaptive monitoring, and dependable AI services operating across distributed infrastructure.

## Scope and responsible interpretation

Drift signals do not automatically prove that a model is unsafe or incorrect. Detectors can produce false alarms, resource measurements can become stale, and placement decisions can carry privacy and reliability consequences. A production implementation needs explicit policies, secure telemetry, confidence thresholds, audit trails, and safe fallback behavior.

Readers should consult the paper, if accepted and published, for the exact method, assumptions, evaluation design, and results.

## Publication status and citation

As of August 6, 2026, this paper is **submitted**, not formally accepted. The workshop lists August 11, 2026 as its acceptance-notification date. No DOI, page numbers, volume, ISBN, or IEEE Xplore record should be inferred before formal publication metadata is assigned.

Current reference:

> N. Mani, X. Li, S. Weerasinghe, and T. Cerny, “SENTRY: Drift-Aware Validation Placement for AI Services in the Cloud-Edge Continuum.” Submitted to the Cloud-Edge Continuum Workshop (CEC 2026), held with IEEE ICNP 2026.

Visit the [CEC 2026 workshop site](https://cecf.github.io/cec26/) for the workshop schedule and current publication information.

## Related research on this weblog

- Browse the complete [Nariman Mani publications list](/publications).
- Explore [multi-agent change impact analysis and test optimization](/posts/multi-agent-change-impact-analysis-test-optimization).
- Read about [layers of AI, from classical to agentic systems](/posts/layers-of-ai-from-classical-to-agentic).


An overview of drift-aware validation placement for AI services operating across cloud and edge environments.


# Secure AI Orchestration for Digital Health: Private Analytics and Adaptive Group Coaching

## Publication at a glance

- **Paper:** *Secure AI orchestration for digital health with differentially private analytics and adaptive group coaching*
- **Authors:** Nariman Mani, Salma Attaranasl, Bahman Sistany, Tomas Cerny
- **Published in:** Smart Health, article 100687 (2026)
- **Persistent record:** [Smart Health 100687](https://scholar.google.com/citations?view_op=view_citation&hl=en&user=yaGTDkcAAAAJ&sortby=pubdate&citation_for_view=yaGTDkcAAAAJ:iH-uZ7U-co4C)
- **Google Scholar:** [View the publication record](https://scholar.google.com/citations?view_op=view_citation&hl=en&user=yaGTDkcAAAAJ&sortby=pubdate&citation_for_view=yaGTDkcAAAAJ:iH-uZ7U-co4C)

## The research problem

Digital coaching platforms need population-level behavioral insight and useful personalization, yet health and lifestyle signals are sensitive. Centralizing raw data or letting loosely governed AI agents exchange it creates privacy, security, and accountability risks.

## The paper's central contribution

The research frames coaching as an orchestrated system rather than a single predictive model. Privacy-preserving analytics, adaptive group formation, coaching workflows, and governance boundaries work together so that useful aggregate signals do not require unrestricted access to individual records.

This distinction matters for both researchers and practitioners: the contribution is a way to reason about the complete system and its decision boundaries, rather than a claim that adding one AI model solves the underlying engineering problem.

## How the approach works

1. **Differentially** private analytics limit what aggregate outputs can reveal about any one participant.
1. **Adaptive** grouping organizes compatible participants while treating privacy constraints as part of the assignment problem.
1. **Secure** orchestration separates responsibilities among analytics, assignment, and coaching components, creating clearer control and audit points.

Together, these elements make the approach easier to study, reproduce, and extend. They also provide clear terms for literature searches and comparisons with related work.

## Why this publication matters

Researchers working on privacy-preserving machine learning, digital therapeutics, health recommender systems, group interventions, or agentic healthcare can use this work as a systems-level reference. It connects privacy mathematics to the operational problem of delivering adaptive coaching at scale.

For practitioners, the larger lesson is to measure the system behavior that matters—not only a model-level score. Reliability, privacy, change over time, human review, and downstream consequences belong in the evaluation.

## Research questions this work can support

This publication is relevant when investigating questions such as:

- How should the core problem be represented so an AI-assisted system retains the right context?
- Which technical and organizational signals provide early, actionable evidence?
- How can automation improve adaptability without concealing failures or weakening safeguards?
- What evaluation design captures effectiveness, safety, privacy, and long-term change?
- Where should a system defer to a developer, architect, coach, clinician, or participant?

## Scope and responsible interpretation

The article should be read as an architecture for balancing utility and privacy, not as permission to treat differential privacy as a complete safety guarantee. Deployment still requires careful privacy-budget selection, threat modeling, clinical validation, fairness analysis, human oversight, and compliance with the rules governing the specific health context.

These boundaries are opportunities for replication, comparison, and follow-on research. Readers should consult the paper itself for its precise method, datasets, experimental setup, and reported results.

## How to cite this paper

> Mani, N., Attaranasl, S., Sistany, B., & Cerny, T. (2026). Secure AI orchestration for digital health with differentially private analytics and adaptive group coaching. Smart Health, 100687.

Use the [Smart Health 100687 publication page](https://scholar.google.com/citations?view_op=view_citation&hl=en&user=yaGTDkcAAAAJ&sortby=pubdate&citation_for_view=yaGTDkcAAAAJ:iH-uZ7U-co4C) to verify the latest bibliographic metadata before submission. You can also find the work through its [Google Scholar record](https://scholar.google.com/citations?view_op=view_citation&hl=en&user=yaGTDkcAAAAJ&sortby=pubdate&citation_for_view=yaGTDkcAAAAJ:iH-uZ7U-co4C). If your research builds on the architecture, method, evaluation, or research framing described here, please cite the original publication rather than this explanatory post.

## Related research on this weblog

- Browse the complete [Nariman Mani publications list](/publications).
- Explore [GraphQL-aware healing with multi-signal learning](/posts/graphql-aware-self-healing-systems).
- For adjacent work, read [adaptive test healing with LLMs and reinforcement learning](/posts/adaptive-test-healing-llm-reinforcement-learning) and [privacy-by-design adaptive group assignment](/posts/privacy-by-design-adaptive-group-assignment).


An accessible guide to secure AI orchestration, differential privacy, and adaptive group coaching in digital health, based on Mani et al. (2026).

Secure AI orchestration for digital health with differentially private analytics and adaptive group coaching


# PRISM-Coach: Privacy-by-Design Adaptive Group Assignment for Lifestyle Coaching at Scale

## Publication at a glance

- **Paper:** *Privacy-by-Design Adaptive Group Assignment for Digital Lifestyle Coaching at Scale*
- **Authors:** Nariman Mani, Salma Attaranasl
- **Published in:** arXiv preprint arXiv:2605.20505 (2026); to appear at IEEE ICHI 2026
- **Persistent record:** [arXiv:2605.20505](https://arxiv.org/abs/2605.20505)
- **Google Scholar:** [View the publication record](https://scholar.google.com/citations?view_op=view_citation&hl=en&user=yaGTDkcAAAAJ&sortby=pubdate&citation_for_view=yaGTDkcAAAAJ:_Qo2XoVZTnwC)

## The research problem

Group coaching can add peer support and improve program reach, but forming useful groups requires behavioral and lifestyle features that participants may not want disclosed. At scale, repeated reassignment also creates a moving privacy and systems-design problem.

## The paper's central contribution

PRISM-Coach treats privacy as a design constraint of adaptive group assignment, not as a later anonymization step. The central question is how a platform can use sufficient signals to form and update useful groups while minimizing exposure and maintaining governable workflows.

This distinction matters for both researchers and practitioners: the contribution is a way to reason about the complete system and its decision boundaries, rather than a claim that adding one AI model solves the underlying engineering problem.

## How the approach works

1. **Privacy-aware** representations restrict the information available to the grouping process.
1. **Adaptive** assignment allows groups to respond to changing engagement and coaching needs instead of remaining static.
1. **A** scalable orchestration layer makes assignment, coaching, and privacy controls explicit system responsibilities.

Together, these elements make the approach easier to study, reproduce, and extend. They also provide clear terms for literature searches and comparisons with related work.

## Why this publication matters

This paper is relevant to researchers studying recommender systems, cohort formation, digital behavior change, privacy-aware clustering, healthcare informatics, and online group interventions. It offers vocabulary for connecting algorithmic assignment quality with data minimization and operational scale.

For practitioners, the larger lesson is to measure the system behavior that matters—not only a model-level score. Reliability, privacy, change over time, human review, and downstream consequences belong in the evaluation.

## Research questions this work can support

This publication is relevant when investigating questions such as:

- How should the core problem be represented so an AI-assisted system retains the right context?
- Which technical and organizational signals provide early, actionable evidence?
- How can automation improve adaptability without concealing failures or weakening safeguards?
- What evaluation design captures effectiveness, safety, privacy, and long-term change?
- Where should a system defer to a developer, architect, coach, clinician, or participant?

## Scope and responsible interpretation

Group compatibility is context dependent, and privacy protection does not automatically ensure fairness, safety, or coaching effectiveness. Future evaluations should examine subgroup outcomes, attrition, adversarial inference, privacy–utility tradeoffs, and how human coaches can challenge an automated assignment.

These boundaries are opportunities for replication, comparison, and follow-on research. Readers should consult the paper itself for its precise method, datasets, experimental setup, and reported results.

## How to cite this paper

> Mani, N., & Attaranasl, S. (2026). Privacy-by-Design Adaptive Group Assignment for Digital Lifestyle Coaching at Scale. arXiv:2605.20505.

Use the [arXiv:2605.20505 publication page](https://arxiv.org/abs/2605.20505) to verify the latest bibliographic metadata before submission. You can also find the work through its [Google Scholar record](https://scholar.google.com/citations?view_op=view_citation&hl=en&user=yaGTDkcAAAAJ&sortby=pubdate&citation_for_view=yaGTDkcAAAAJ:_Qo2XoVZTnwC). If your research builds on the architecture, method, evaluation, or research framing described here, please cite the original publication rather than this explanatory post.

## Related research on this weblog

- Browse the complete [Nariman Mani publications list](/publications).
- Explore [GraphQL-aware healing with multi-signal learning](/posts/graphql-aware-self-healing-systems).
- For adjacent work, read [adaptive test healing with LLMs and reinforcement learning](/posts/adaptive-test-healing-llm-reinforcement-learning) and [privacy-by-design adaptive group assignment](/posts/privacy-by-design-adaptive-group-assignment).


How privacy-by-design group assignment can support scalable, adaptive digital lifestyle coaching without exposing sensitive participant data.

Privacy-by-Design Adaptive Group Assignment for Digital Lifestyle Coaching at Scale


# Key Developer Roles and Organizational Coupling in Microservices: A Longitudinal View

## Publication at a glance

- **Paper:** *Key Developer Roles and Organizational Coupling in Microservices: A Longitudinal Analysis*
- **Authors:** Xiaozhou Li, Nariman Mani, Jose Sosa Rodriguez, Tomas Cerny
- **Published in:** arXiv preprint arXiv:2604.25804 (2026); EASE 2026
- **Persistent record:** [arXiv:2604.25804](https://arxiv.org/abs/2604.25804)
- **Google Scholar:** [View the publication record](https://scholar.google.com/citations?view_op=view_citation&hl=en&user=yaGTDkcAAAAJ&sortby=pubdate&citation_for_view=yaGTDkcAAAAJ:R3hNpaxXUhUC)

## The research problem

Microservices promise independently evolvable services, but repository boundaries do not guarantee independent teams. A small number of developers may quietly connect many services through their contribution and coordination patterns, creating knowledge concentration and architectural risk.

## The paper's central contribution

The paper studies these relationships longitudinally. Instead of taking one snapshot of ownership, it asks how key developer roles and cross-service organizational coupling evolve over time and what those changes reveal about the actual architecture-producing organization.

This distinction matters for both researchers and practitioners: the contribution is a way to reason about the complete system and its decision boundaries, rather than a claim that adding one AI model solves the underlying engineering problem.

## How the approach works

1. **Contribution** histories expose who repeatedly works within or across service boundaries.
1. **Longitudinal** analysis distinguishes persistent coupling from a brief migration, incident, or release-driven collaboration.
1. **Role** and coupling measures provide socio-technical signals that can be compared with intended service ownership.

Together, these elements make the approach easier to study, reproduce, and extend. They also provide clear terms for literature searches and comparisons with related work.

## Why this publication matters

Researchers and practitioners can connect this work to bus-factor analysis, developer networks, code ownership, socio-technical congruence, and microservice maintainability. The paper is particularly useful when repository topology says services are separate but delivery work suggests otherwise.

For practitioners, the larger lesson is to measure the system behavior that matters—not only a model-level score. Reliability, privacy, change over time, human review, and downstream consequences belong in the evaluation.

## Research questions this work can support

This publication is relevant when investigating questions such as:

- How should the core problem be represented so an AI-assisted system retains the right context?
- Which technical and organizational signals provide early, actionable evidence?
- How can automation improve adaptability without concealing failures or weakening safeguards?
- What evaluation design captures effectiveness, safety, privacy, and long-term change?
- Where should a system defer to a developer, architect, coach, clinician, or participant?

## Scope and responsible interpretation

Repository activity is an imperfect proxy for communication, responsibility, or expertise. Bots, pair programming, reviews, platform teams, and organizational changes can alter the interpretation. Coupling indicators should initiate architectural inquiry, not rank individual developers or serve as employee-performance metrics.

These boundaries are opportunities for replication, comparison, and follow-on research. Readers should consult the paper itself for its precise method, datasets, experimental setup, and reported results.

## How to cite this paper

> Li, X., Mani, N., Sosa Rodriguez, J., & Cerny, T. (2026). Key Developer Roles and Organizational Coupling in Microservices: A Longitudinal Analysis. arXiv:2604.25804.

Use the [arXiv:2604.25804 publication page](https://arxiv.org/abs/2604.25804) to verify the latest bibliographic metadata before submission. You can also find the work through its [Google Scholar record](https://scholar.google.com/citations?view_op=view_citation&hl=en&user=yaGTDkcAAAAJ&sortby=pubdate&citation_for_view=yaGTDkcAAAAJ:R3hNpaxXUhUC). If your research builds on the architecture, method, evaluation, or research framing described here, please cite the original publication rather than this explanatory post.

## Related research on this weblog

- Browse the complete [Nariman Mani publications list](/publications).
- Explore [GraphQL-aware healing with multi-signal learning](/posts/graphql-aware-self-healing-systems).
- For adjacent work, read [adaptive test healing with LLMs and reinforcement learning](/posts/adaptive-test-healing-llm-reinforcement-learning) and [privacy-by-design adaptive group assignment](/posts/privacy-by-design-adaptive-group-assignment).


A research guide to key developer roles and organizational coupling in microservices, with implications for architecture governance and socio-technical risk.