A detailed project archive across research, products, and creative systems.

This system was built to guide users through the compilation of ORA requests for BAZL. The core idea is not just to attach a large model to a user interface, but to engineer a retrieval pipeline where chunks of source text are embedded, indexed, scored, re-ranked, and assembled into a context package before the model answers.

Candidate chunks are extracted from the indexed document corpus starting from a natural-language query. They are then filtered and re-ranked through dense retrieval, keyword matching, and reranking stages to create a trustworthy knowledge base for the model. The project also investigated the use of knowledge graphs as an additional retrieval improvement layer.

• Distributed architecture with Node.js frontend, Flask backend orchestration, Ollama-based model execution, and PostgreSQL persistence.
• Designed for regulatory text, where precision, grounding, and traceability matter much more than generic fluency.
• Open-source codebase intended as a practical compliance support system rather than a demo chatbot.

The framework binds an aerodynamic simulator to an open-source PID-based drone controller and evaluates alternative control strategies under iterative optimization. The goal is not only to improve performance, but also to reduce disturbance, battery consumption, trajectory aggressiveness, and rotor loudness.

The optimization layer compares multiple approaches, including Bayesian optimization, reinforcement learning, and metaheuristic methods, and studies how they perform under the same physical simulation setup. This makes the project both a control-engineering investigation and a methodological benchmark.

• Unified simulation pipeline combining physical models, control logic, and acoustic proxy modeling.
• Optimization campaign focused on public disturbance reduction and flight efficiency.
• Research direction that naturally connects to drone acoustics and psychoacoustic annoyance modeling.

This work supported a PhD-level research effort aimed at building a surrogate architecture capable of predicting aerodynamic quantities over very large 3D CFD meshes. The main challenge was dimensionality reduction without destroying physically relevant structure, especially in nonlinear or high-gradient regions.

A graph-based autoencoder emerged as the strongest architectural choice. It enabled mesh-aware encoding and reconstruction while preserving useful geometric and aerodynamic relations across nodes. The project also evolved into spatio-temporal forecasting work, extending the surrogate idea from steady predictions to dynamic pressure-field evolution.

• Graph encoders and decoders operating directly on aerodynamic mesh structures.
• Reduced-space and reconstruction strategies for large-scale scientific simulation data.
• Research direction with strong evidence of drastic compute reduction relative to standard CFD.

In many simulation workflows, the same input can produce different outcomes depending on the fidelity of the generating model. This project addresses that mismatch directly by building a stochastic neural architecture that can learn from multiple fidelity levels rather than treating them as isolated datasets.

The resulting model estimates both predictions and uncertainty bands, making it more suitable for engineering decision support where the confidence of the estimate matters as much as the estimate itself. Validation was performed across two aerodynamic case studies.

• Deep-learning-based fusion of low-, medium-, and high-fidelity simulation sources.
• Uncertainty-aware output rather than purely deterministic regression.
• Transfer-learning logic used to extract more value from scarce high-fidelity data.

The system was structured as a layered architecture rather than a single monolithic application. Sensor data, reasoning modules, machine-learning components, and service interfaces were organized to support both predictive behavior and interpretable higher-level decision making.

Particular attention was given to security and performance evaluation, which is critical for distributed smart-city systems operating on heterogeneous signals and constrained infrastructure.