Projects

macOS menu bar app that runs a local Whisper inference server as a managed subprocess, captures audio via global hotkey, and injects transcribed text into any app — entirely on-device, no cloud, no latency from network.

• —5 independent components coordinated: ML server, keyboard tap, audio pipeline, inference API, clipboard injection
• —Near-zero CPU at idle · <100MB resident memory
• —Full test coverage across all major components
• —Video demo available on the detail page — app demo and setup walkthrough

AI-driven decision support for commodity traders and buyers — turning raw LME price data and GDELT news signals into actionable buy / hedge / wait recommendations. Built at the CHEFTREFF FLS26 Hackathon with a team of 6.

• —Decision cockpit focused entirely on high-stakes trading decisions — buy now vs. wait, hedge vs. stay exposed
• —Explainability first — logistic regression, XGBoost, and decision trees so traders understand the "why"
• —Data edge: integrated a public metal production plant thermal index dataset for stronger signal
• —Offline ML pipeline → snapshot API → Streamlit frontend; deployable via Docker

A recipe app built to demonstrate production engineering practices — the domain is intentionally simple so the focus stays on the infrastructure and delivery.

• —CI pipeline on every PR: Jest unit tests, Angular build, and Playwright e2e — all running in Docker
• —Kubernetes with K3d — Kustomize overlays for env config, Traefik and Nginx ingress both supported
• —3 Docker Compose files: dev (bind mounts), prod simulation, and CI (clean builds)

Master's Thesis — automated pipeline that uses open-source LLMs to generate Question-Answer-SPARQL triplets from DBLP Knowledge Graph triples in a single prompt, without fine-tuning.

• —87.38% accuracy — 10,704 correct QASP triplets generated from 12,250 total using Qwen2.5-Coder-32B
• —6 models benchmarked across 6 triple format types, 3 metadata levels, and 0/1-shot settings
• —Key finding: less structural metadata yields higher accuracy — more context confuses rather than helps

Create and join chat rooms, send messages, and see live typing indicators — no account required. Built to explore bidirectional event-driven communication with Socket.IO.

• —Live typing indicators and join/leave notifications via dedicated socket events
• —Last 50 messages loaded on join — persisted in MongoDB, not just in-memory
• —Duplicate username prevention per room · active user list synced across all members
• —Built-in onboarding tour — coach marks tutorial guides new users through the UI on first visit

Benchmarked various GAN architectures, focusing on mathematical foundations and advancing from Vanilla GANs to Wasserstein GANs for improved stability.

Solved the Bus-Routing-Problem for KVG (regional transit operator) using linear optimization techniques. Implemented efficient routing algorithms for real-world logistics.

Analysis and presentation of audio signal processing in OpenAI Whisper Model. Focused on Audio processing and Beam-Decoding for speech-to-text conversion.

Probabilistic modelling project implementing Bayesian updating of building vulnerability curves for hurricane damage assessment using a Zero-One Inflated Beta (ZOIB) distribution — extended with a CNN-based damage classification pipeline on satellite imagery.

• —ZOIB likelihood model with NUTS sampler — 3000 samples, 3 chains, posterior narrows uncertainty bands significantly
• —Hurricane Dorian (2019) observations: 732 buildings from 15 YouTube videos classified by 3 experts
• —Sequential CNN on xBD satellite dataset — 75.7% accuracy across 4 damage classes with minimal overfitting

Physics-based 2D UAV simulation that autonomously chases a target balloon using equations of motion and cascaded PID control. Built in C++/SFML with gain tuning done in MATLAB Simulink.

• —Rigid body physics: dual propeller thrust, gravity, tilt angle — integrated per frame
• —Two cascaded PID loops: altitude (Y error → thrust magnitude) and attitude (X error → thrust differential)
• —Gains tuned in Simulink, ported to C++ — position ramps to target in ~3.5s with clean second-order response

Research presentation on SAT (Structure Attribute Transformer) — a model for learning on graphs where nodes have missing or entirely absent feature vectors, combining dual VAE pipelines with adversarial distribution matching.

• —Two parallel VAE pipelines over node attributes X and graph structure A — tied via PSAM and ADM
• —Adversarial Distribution Matching replaces KL divergence — shared discriminator aligns both latent spaces
• —Restores missing node attributes and predicts missing links after training

Simulation of Lost Target Search using Modified Particle Swarm Optimization with PID-based flight stability. Multi-language implementation for research purposes.

Sensorless BLDC motor driver using an Arduino Uno. Uses back-EMF sensing via the AVR's internal analog comparator to estimate rotor position and perform six-step commutation — no hall effect sensors required. Drives the motor through a three-phase MOSFET gate driver/inverter circuit.

As part of the TopOfStack Software team, contributed to the large-scale conversion of MyComparateur from Adobe Flash/AIR (MXML and ActionScript) to Vue.js and Node.js. Converted MXML UI layouts to Vue.js components and migrated ActionScript business logic to Node.js backend for a modern web implementation.

A chat application focused on authorization during login, showcasing Google Cloud database (Firebase) and Vue Js routing, props, and authorization features. Used Materialize CSS for styling.

Led the development of an institutional framework for continuous learning within Ma-Co, a maritime training organisation in Hamburg. Applied sensemaking theory to map organisational change, supported by interview and survey analysis.

• —5-step knowledge management pipeline: cue extraction → necessary condition analysis → working model → implementation modules
• —Sensemaking framework: trigger-processing-results cycle applied to organisational change
• —Lean philosophy, cross-functional collaboration, and KPI-based performance monitoring

Research presentation on the evolving landscape of data privacy — covering GDPR, emerging technology risks (AI, quantum computing, blockchain), individual protection tools, and the ethical gap between what can be done technically and what should be done responsibly.

• —GDPR as global gold standard — analysis of global enforcement variance and inspired legislation
• —Emerging risks: AI re-identification & bias, quantum decryption, blockchain compliance tension
• —eTA (Ethical Technology Assessment) framework — assess before deploying, not after

A salon management platform developed as part of a team at Encoder IT Limited, featuring staff management, task scheduling, calendar functionality, e-commerce integration, inventory management, user rating system, and service booking system.