My student Estela Mora Barba disserted her Internet of Things Master thesis

Created in July 16, 2026 by Néstor Lucas Martínez

2026   ·   master-thesis   MLOps   edge-computing   IoT   trustworthy-AI   honors   ·   news

🎓 Outstanding Master’s Thesis Defense On Thursday, July 16, 2026, my student Estela Mora Barba successfully defended her Master’s Thesis titled “An AI Inference and Automated Deployment Framework for IoT across the Edge-Cloud Continuum.” Her work received the highest possible grade awarded that day, a 9.5/10, from the evaluation committee.

Context & Motivation

Deploying AI on the edge, close to where data is actually generated, is one of the defining engineering challenges of the current wave of IoT systems. It is also one of the hardest to get right in practice, since it means reconciling constrained hardware, unreliable connectivity, and the operational demands of running machine learning models reliably outside the comfort of a data center. Estela’s thesis takes on this challenge directly, focusing on the “Eco-Friendly Multimodal Mobility” use case within the European research initiative NexTArc (funded under the Horizon Europe programme) and aligned with the Chips JU roadmap. Her work contributes concretely to European digital sovereignty and to the practical realization of Trustworthy AI principles (confidentiality, integrity, and availability) in real deployment conditions, not just as design goals on paper.

System Architecture & Key Contributions

To tackle hardware fragmentation and the tight resource budgets typical of edge devices, Estela designed and implemented AURA (AI Unified Runtime & Analytics), an open, distributed, edge to cloud MLOps platform:

  • Platform Abstraction Layer (PAL): a hardware agnostic runtime core running directly on IoT edge devices that decouples AI inference from the specifics of the underlying hardware, enabling real time local object detection (YOLO) regardless of which accelerator sits underneath. In this context, inference refers to the process of running an already trained AI model on new input data to produce a result, precisely the definition formalized in ISO/IEC 22989:2022, the international standard for AI concepts and terminology, which builds on the broader IT vocabulary established in ISO/IEC 2382. Estela’s inference loop runs on a strict 100 millisecond cycle, keeping the system well within real time operating bounds.
  • Automated MLOps Pipeline: end to end orchestration covering training and retraining, automated hardware specific model compilation and quantization (including INT8 optimization through Sony’s Model Compression Tool for the IMX500 target), and Over The Air (OTA) atomic releases with dynamic rollback.
  • Technology Stack: the cloud control plane relies on FastAPI, PostgreSQL, MongoDB, Redis, and MinIO, complemented by a Next.js and React dashboard. Edge to cloud interactions rely on gRPC for internal microservices, MQTT for edge commands and telemetry, and Prometheus for monitoring.
  • Validated Edge Backends: the framework was proven across diverse real world targets, including the Raspberry Pi 5 CPU, Hailo-8/8L PCIe accelerators, and the Raspberry Pi AI Camera (Sony IMX500). Notably, the Hailo-8L configuration required just 588.8 MB of RAM under load, about 56% less than the CPU only baseline, showing how much headroom hardware acceleration frees up for other edge tasks.

Validation & Impact

What really stands out is the depth of the verification and validation work behind AURA. Estela built her V&V campaign around the concepts and test level taxonomy of the ISO/IEC/IEEE 29119:2020 software testing standard and IEEE 1012, structuring her evaluation into clearly separated functional, non functional, and system level test suites. Across both Software in the Loop (SiL) and Hardware in the Loop (HiL) configurations, and under two representative deployment scenarios (single tenant and multi device), she exercised the complete AURA pipeline end to end: device registration, asynchronous model compilation, OTA delivery with SHA-256 integrity checks, on device YOLO inference, MQTT telemetry, and dashboard synchronization. The results speak for themselves: consistent 100% success rates across thousands of gRPC calls, database transactions, and deployment cycles, with zero critical failures, even under network disruption and concurrent load.

Beyond the numbers, the thesis lands squarely on the strategic goals of the NexTArc project. The platform’s edge acceleration and quantization pipeline directly targets NexTArc’s real time and resource efficiency objectives, its automated OTA orchestration addresses the project’s operational cost reduction goals by letting a single administrator manage large fleets of distributed nodes without physical intervention, and its reliance on affordable commercial hardware supports the reduction of capital expenditure as well. On top of that, the “Privacy by Design” approach, transmitting only aggregated telemetry and never raw video, and the atomic rollback mechanism in the OTA handler, give the platform the kind of resilience and trustworthiness that NexTArc’s demonstrators are meant to showcase. That is exactly the kind of empirical rigor and strategic alignment a real world MLOps platform for edge AIoT needs to earn before anyone should trust it in production.

One of the evaluation committee members was so impressed with the quality and technical depth of the presentation that they explicitly recommended expanding and adapting the work for submission to a top tier JCR journal, an initiative that was already planned and is currently underway!

This work paves the way for Estela’s continued research within the PhD program in Systems and Services Engineering for the Information Society (DISSSI). Congratulations, Estela, on this well deserved success!