KIconnect AI Platform - Full-Stack Development
Build a production-ready, multi-tenant SaaS platform from scratch with real-time streaming capabilities, complex user authentication via enterprise SSO, encrypted data storage, and modern reactive frontend - all while maintaining high availability and EU compliance.
- Backend Architecture: Designed and implemented ASP.NET Core backend with clean architecture principles, dependency injection, and CQRS patterns
- Real-time Communication: Built SignalR hubs for bi-directional streaming of LLM responses with backpressure handling and automatic reconnection using Polly retry policies
- Modern Frontend: Developed Vue 3 + Composition API admin dashboard with Inertia.js for SPA-like experience without API overhead, bundled with Vite for optimal performance
- Security & Authentication: Integrated Shibboleth SSO for enterprise single sign-on, implemented MongoDB Client-Side Field Level Encryption (CSFLE) for sensitive data
- Multi-tenancy: Designed tenant isolation at database level with per-tenant quotas, rate limiting, and usage tracking
- DevOps: Deployed on Nomad orchestrator with Traefik reverse proxy, implemented background job processing with Hangfire
- Model Management: Built admin interface for managing LLM deployments, reasoning effort configuration, and model version control
✓ 99.9% uptime with automatic failover and circuit breaker patterns
✓ Reduced administrative overhead with self-service tenant management
✓ Ensured EU data compliance and security with field-level encryption
TalkEcho - Real-time Meeting Subtitle & Translation
Invisible real-time meeting subtitles and translation assistant. TalkEcho runs as a translucent desktop overlay that captures both system audio and microphone input, streams them to speech-to-text and LLM providers, and returns bilingual captions without disrupting the meeting flow.
- Real-time audio capture from system and microphone
- Live speech-to-text transcription with multiple provider support
- On-the-fly translation to multiple languages using LLMs
- Transparent desktop overlay that doesn't interfere with applications
- Configurable caption position, size, and appearance
GitHub Repository (Forked from Pluely)
CPM Remote Web Application
Maintain and enhance a complex web application that allows users to remotely test autonomous vehicle algorithms in real-time. The system must manage session requests, run backend simulations, visualize experiments on frontend, and support algorithmic competitions with stage-based updates.
- Session Management: Implemented robust session queue system handling concurrent user requests for limited simulation resources
- Real-time Visualization: Developed real-time data visualization dashboards using Angular to display simulation results and vehicle telemetry
- CI/CD Pipeline: Enhanced GitLab CI/CD pipelines for automated testing, building, and deployment of both frontend and backend components
- Git Submodule Architecture: Managed complex multi-repository structure with git submodules for modular algorithm integration
- Competition Platform: Built stage-based competition system with leaderboards, submission management, and automated evaluation
✓ Supported international algorithmic competitions with 100+ participants
✓ Reduced development cycle time with improved CI/CD automation
WebOCD: Dynamic Graph Community Detection Platform
Extend the WebOCD (Web-based Overlapping Community Detection) platform to support dynamic graphs that change over time, implementing efficient algorithms for detecting overlapping communities as the graph evolves.
- Designed and implemented dynamic graph data structures supporting temporal evolution of networks
- Integrated COPRA (Community Overlap Propagation Algorithm) for overlapping community detection
- Built RESTful APIs for graph manipulation, algorithm execution, and result visualization
- Optimized algorithm performance for large-scale graphs with thousands of nodes and edges
- Developed frontend visualizations for temporal network analysis
✓ Provided researchers with tool for studying dynamic network phenomena
✓ Bachelor thesis grade: 1.7 (Good)
Thesis: "Algorithmic Approaches to Overlapping Community Detection - Community Overlap Propagation Algorithm (COPRA)"
Bank Web Application - FinTech Data Analytics
Develop a full-stack banking web application from scratch with modern architecture, providing data analytics features for financial institutions while ensuring secure transaction handling.
- Built complete MEAN stack application (MongoDB, Express/Node.js, Angular)
- Designed RESTful API architecture for account management, transactions, and analytics
- Implemented secure authentication and authorization with JWT tokens
- Created interactive data visualization dashboards for financial analytics
- Containerized application with Docker for consistent deployment
✓ Delivered production-ready architecture patterns for financial applications
✓ Course grade: 1.3 (Very Good)
Personal Website & Portfolio
Responsive personal website showcasing professional experience, projects, and skills. Automatically builds and deploys CVs from Typst source files using GitHub Actions.
- Fully responsive design with mobile-first approach
- Categorized project showcase by professional role
- Automated CV generation pipeline (Typst → PDF → GitHub Pages)
- Integrated blog with Giscus comments
- SEO-optimized and accessible design
Other Web & Mobile Projects
Tech: TypeScript | Snore detection application using audio analysis
Tech: JavaScript | Audio recording and processing tool
Self-hosted Matrix & Bots
A tiny two-person universe for me and my partner. Self-hosted Synapse (rickandzoey.com) + Element Web with end-to-end encrypted chats. A couple of utility bots deliver alerts for daily life and side projects.
Headless Playwright bot that watches Bürgeramt/SuperC appointment slots and pings our Matrix room when something opens. Handles cookie banners and tricky modal dialogs, and can attach screenshots or heatmaps of hot times.
Daily digest for BTC, S&P 500, and Gold. Computes rolling drawdown/surge, triggers threshold alerts with context, and uploads charts to our Matrix room. Keeps greed and fear in check, one ping at a time.
Lightweight quality checks for retrieval‑augmented generation pipelines. Runs evaluations, records telemetry, and posts a Matrix alert if metrics drop—so my AI doesn’t silently drift.