Theses

If you wish to propose your own topic in a related field, please feel free to do so and contact us!

Open theses topics for the 2026/2027 study year

List of Supervisors

• Pelle Jakovits
• Vinayak Khavasi
• Andres Namm
• Anjish Panta

Synthetic IoT data generator for large-scale IoT Device simulation (M) (Pelle Jakovits)

The student should evaluate existing open source tools for generating real-like IoT data based on existing captured data traces and “play it back” to simulate real data in an IoT network. The goal is to design and create a solution which processes existing data traces and can generate similar behaving data stream with high-volume and frequency. It should also support customizing the generated data stream, including volume and frequency, structure (like ratios between the different types of sub-streams), randomizing certain fields of records. Such tool could be used to test and evaluate the performance of open source IoT platforms.

Reliability and performance of Open Source IoT platforms for Smart Cities (B, M) (Pelle Jakovits)

The goal of this topic is to analyse the reliability and performance of open source IoT platforms focusing on solutions that support large use cases (e.g not home automation but rather Smart City, devices deployed over large geographical locations). The main aspects to focus on are the performance, stability and the scope of features (e.g whether they support Video processing in the Edge). The main questions to answer are:

• Which OS solutions in the market are the most suitable for supporting large Smart City use cases?
• Are open source IoT frameworks mature enough for production systems, and are they feasible alternatives to cloud-managed subscription-based IoT platforms?
• How well the Open source IoT platforms support Edge Video processing?

Root Cause Analysis or IoT device network failures (B, M) (Pelle Jakovits)

The goal is to find patterns in data that describe problematic IoT devices. The initial use case is water metering devices that send data packets and have a data transmission credit – which represents the quality of data transmission connectivity. The presence of a problem is indicated by a continuous decrease in the credit for sending data or reaching zero. Some of the questions to investigate are: Based on all the data points, what is the pattern of the problem device? What parameters define the problematic devices?

Automated Grader for the “Development of Web Services and Distributed Systems” course (B/M) (Pelle Jakovits)

A grading system based on automated testing or monitoring frameworks needs to be designed and implemented for usage in the Development of web services and distributed systems course for students and course organizers. The thesis might also include redesigning course lab materials and tasks. An Estonian speaker is expected as the cours eis taught in Estonian.

Cost prediction & optimization tool for Serverless applications (B, M) (Pelle Jakovits)

Serverless (or FaaS) applications enable the Event based execution model for Cloud services where you do not need to pay for the idle run time but instead pay for how long it takes for functions to be executed and how much memory is allocated. By optimizing memory consumption or runtime it is possible to reduce the cost of Serverless applications. Investigate what characteristics affect the cost of Serverless applications in the Cloud the most, compare to non-Serverless approaches and find ways to reduce the cost of Serverless applications. The goal of this thesis is to create a web based tool for predicting the cost of serverless application deployments.

Tool for estimating the energy consumption of buildings in Smart Cities (B, M) (Pelle Jakovits)

The goal of this thesis is to collect and enhance Tartu City building metadata from different data sources, like Estonian state registries, to estimate how much energy is consumed by the building. Taking into account the type of construction, materials, age of the building, and registered heating and cooling systems inside the building or its apartments. This topic can build upon an earlier thesis that designed a working solution for estimating the energy production of Smart City buildings. We also have data available from a number of Tartu City Government buildings and SmartEnCity project buildings. Output of the thesis should be a tool or software that can be used to visualize city buildings, their metadata and energy estimations.

Automated meta-data collector and enhancer for Smart City building energy profiling (B, M) (Pelle Jakovits)

The goal of the thesis is to create a solution which would automatically extract metadata of buildings for a specific (Estonian) city from multiple different data sources (building registries, 3D model data of city buildings, etc. ) and registries and prepare this data for building energy profiling simulations (focusing on energy production, consumption and balance). The results can be used as an input to different building energy simulation tools with the goal to perform city-scale building energy analysis and optimization.

Web interface for Smart City data platform (B) (Pelle Jakovits)

The goal is to create a Web frontend for a Smart City API based platform which is used to store large amounts of data sent by Smart City and IoT devices. Investigate existing solutions, collect and document requirements, design a working solution, and deploy it in university Cloud.

Web interface for Dagster data pipelines (B) (Pelle Jakovits)

The goal is to create a Web frontend for defining data pipelines in Dagster. For example between external APIs and an API based Smart City database which is used to store large amounts of data sent by Smart City and IoT devices. Students should investigate existing solutions, collect and document requirements, design a working solution, and deploy it in university Cloud. This topic can build upon an earlier thesis that designed a working solution for implementing and automating Dagster based data integration pipelines.

Camera-based sensing in Smart Cities (M) (Pelle Jakovits)

A large number of sensors have and are being deployed in IoT networks and Smart Cities. However, it is not be feasible to deploy a new sensor every time a new type of observation/data needs to be detected. It is sometimes more feasible to use more general types of devices – like Visual and Thermal Cameras – to detect different types of events and measure data. Investigate what types of Smart City events can be detected, how to ensure data privacy when processing camera feeds, what are the most critical challenges in using cameras for the detection of events in Smart Cities. Design a prototype solution for camera or thermal camera based device that processes the visual stream locally with data privacy guarantees.

Edge Intelligence Observability: Lightweight and Hybrid Monitoring Approaches  (M) (Vinayak Khavasi)

Observability—the ability to measure the internal state of a system by observing external outputs—is a cornerstone of reliable distributed computing. In the context of edge intelligence, observability becomes uniquely challenging due to the constrained resources, intermittent connectivity, and real-time requirements of edge devices. Traditional observability and MLOps tools such as MLFlow, EvidentlyAI, and Alibi Detect offer comprehensive functionality but are often too resource-heavy for microcontrollers or lightweight IoT devices. This thesis explores the design and evaluation of lightweight observability methods tailored for edge AI, focusing on hybrid architectures that combine on-device minimal monitoring with edge gateway/local aggregation strategies. By synthesizing existing open-source tools with custom statistical metrics (e.g., mean, variance, categorical counts, error rates) and streaming libraries such as River, the goal is to propose an efficient and practical framework that ensures reliable performance monitoring, drift detection, and anomaly identification in resource-constrained environments.

Industrial IoT Edge AI Observability: Predictive maintenance via observability of embedded AI models in manufacturing equipment.  (M) (Vinayak Khavasi)

The benefits of observability in AI models extend to enhanced decision-making capabilities, as it provides insights into system performance and potential areas for improvement, fostering a proactive maintenance culture. Additionally, observability facilitates the identification of anomalous patterns in equipment behavior, enabling timely interventions that can prevent costly breakdowns and extend asset lifespan. This proactive approach not only minimizes unplanned downtime but also optimizes resource allocation, ultimately leading to significant cost savings and improved operational efficiency in manufacturing environments. The integration of AI models into manufacturing processes not only enhances predictive maintenance but also supports the broader goals of Industry 4.0 by improving efficiency and connectivity across systems.

Observability in Edge AI vision models for autonomous vehicles  (M) (Vinayak Khavasi)

Investigate observability in Edge AI vision models for autonomous vehicles, with a focus on ensuring safety and reliability through runtime monitoring. The work should aim to identify effective observability metrics (e.g., drift detection, uncertainty estimation, anomaly monitoring) and develop lightweight monitoring mechanisms suitable for resource-constrained edge devices. By exploring trade-offs between performance, latency, and safety, the study should propose an observability framework/architecture that can detect critical failures, adapt to real-world distribution shifts (e.g., weather, lighting, sensor degradation), and provide interpretable safety signals for both system operators and regulators.

Unified Observability Framework for Edge AI  (M) (Vinayak Khavasi)

Unified observability for edge AI represents a critical convergence of distributed systems monitoring, machine learning operations (MLOps), and edge computing infrastructure. Synthesize current research and industry practices to provide a comprehensive framework for implementing observability solutions that span the edge-to-cloud continuum.
Explain how an unified approach combines traditional observability pillars (metrics, logs, traces) with AI-specific monitoring requirements (model performance, data drift, inference quality) and edge-specific constraints (resource limitations, network intermittency, privacy requirements).

Log fingerprint recognition using computer vision methods (M) (Andres Namm)

Objective: to test and compare existing computer vision methods for detecting unique surface patterns (cross-sectional annual rings, cracks, bark marks) of logs. Public datasets such as Biomtrace (photos of cut edges of logs) and ATECH2024 (image processing and matching models) are used. Methods:

• segmentation of cross-sectional surfaces,
• feature extraction and comparison (e.g. SuperPoint, LightGlue),
• testing on photos and data of both individual logs and log stacks,
• evaluation of results in terms of accuracy and robustness.

Result: the work provides an overview of how well existing methods work in determining the individual identity of logs both individually and in stacks, and how suitable they may be in supply chain traceability solutions.

Root-Cause Analysis of Latency and Data-Pipeline Failures in Industrial Edge AI (Vinayak Khavasi)

Industrial Edge Intelligence also referred as Edge AI systems used for visual inspection or predictive maintenance depend on coordinated sensor, preprocessing, model-serving, hardware, and application components. This thesis will investigate whether (distributed/edge) observability can identify whether an AI-service degradation originates from the data pipeline, edge infrastructure, model execution, or downstream application. The scope will be limited to one industrial use case, one representative model, and two failure classes: inference-latency degradation and data-pipeline quality failure. A prototype will collect lightweight metrics and traces for data freshness, preprocessing time, queueing, model execution time, resource utilization, and model version. These signals will be correlated to produce an explainable root-cause diagnosis rather than only a generic health alert. The evaluation will measure detection accuracy, diagnosis

Low-Overhead Monitoring of TinyML Crop-Health Models Under Seasonal Drift (Vinayak Khavasi)

Agricultural Edge AI also referred as Edge Intelligence systems often operate on low-power devices with limited memory, intermittent connectivity, and changing environmental conditions. This thesis will study how a TinyML crop-health or soil-monitoring model can be monitored without consuming excessive energy, memory, or communication bandwidth. The scope will be limited to one sensing task, one small device platform, and two monitoring objectives: resource overhead and seasonal or environmental drift. The proposed monitor will use adaptive sampling and local aggregation for energy, memory, sensor-quality, confidence, and data-distribution signals. Weather, season, location, or soil context will be used to distinguish expected environmental variation from potentially harmful model degradation. The evaluation will compare monitoring overhead, drift-detection quality, false-alert rate, and model latency against a monitoring-disabled baseline. The expected contribution is a practical low-overhead observability strategy for TinyML deployments rather than a new crop-detection algorithm.

Context-Aware Drift Detection for Edge-Based Urban Traffic Analytics (Vinayak Khavasi)

Smart-city traffic analytics rely on distributed edge cameras and sensors that operate under changing weather, lighting, traffic, event, and infrastructure conditions. This thesis will investigate whether contextual information can reduce false alarms and improve the detection of harmful drift in one edge-based traffic-analytics model. The scope will be limited to one task, such as vehicle counting or congestion classification, and a small set of context variables such as time of day, weather, traffic volume, and planned events. The monitoring system will compare conventional feature-distribution alerts with context-aware alerts that incorporate external conditions and delayed performance feedback. It will also record inference latency, camera or pipeline health, model version, and edge-node load to distinguish model drift from infrastructure problems. The evaluation will measure drift-detection precision, false-alert rate, detection delay, and resource overhead under controlled traffic scenarios. The expected contribution is an experimentally validated context-aware monitoring method for a distributed smart-city Edge AI service, not a complete city-scale platform.

A Configurable Time-Series Workload Generator for Emulating SD Card Wear in Edge IoT Devices (Anjish Panta)

The goal of this topic is to build an open-source tool that simulates realistic IoT storage workloads on Raspberry Pi devices. The tool generates mock sensor data, writes it to a time-series database on the SD card, and runs Grafana-style queries on it, just like a real IoT system. The main aspects to focus on are realistic data generation, easy configuration, and repeatable experiments. The tool will be tested on a Raspberry Pi cluster, and SD card wear will be monitored with Prometheus and Grafana. The main questions to answer are:

• How can realistic IoT storage workloads be simulated in a simple and repeatable way?
• How do data rate, batch size, and query frequency affect how much data is written to the SD card?
• Which time-series database is the most SD card friendly for long-running IoT devices?

Automatic Position Adjustment System for Temperature Monitoring with a Thermal Camera (Anjish Panta)

The goal of this topic is to improve thermal camera software that measures the temperature of SD cards in a Raspberry Pi cluster. Currently, temperatures are read from fixed positions in the image, so even a small camera movement gives wrong readings. The new system uses a fixed reference point in the camera’s view, detects how much it has moved, and adjusts the measurement positions automatically. The main aspects to focus on are accuracy, stability, and ease of use. The main questions to answer are:

• How reliably can a fixed reference point be detected in a thermal image?
• How much does automatic adjustment improve accuracy compared with fixed positions?
• How much camera movement can the system handle while still giving correct readings?