Michal Ondrejka

Conduit — RAG + MCP server

Conduit is a RAG + MCP server that gives any MCP client semantic search over your data — work items, code, docs and test results from any JSON-over-HTTP API — plus a persistent Experience store for remembering facts, preferences and past decisions across sessions. It ships as a single Go binary backed by Qdrant and any OpenAI-compatible embedding endpoint.

It exposes 9 MCP search tools covering work items, requirements, source and test code, builds, commits and documentation. Sources are provider-agnostic: any JSON API is connected purely through configuration (URL, auth, field mappings, pagination), so Azure DevOps, Jira and GitHub need no provider-specific code. A web UI manages sources, credentials and syncs; secrets are stored Fernet-encrypted and referenced by name; and indexing is transactional, so a failed sync leaves existing data intact.

AR Sandbox with water simulation using hand gestures

The goal of this work was the design and implementation of an interactive learning system that allows the user to explore and understand weather conditions in real time. The implementation is based on the existing AR Sandbox solution. I researched the Lattice Boltzmann Method to simulate water, but in the end decided to use Smoothed Particle Hydrodynamics: the liquid is represented using sphere particles and then visualized as an object similar to liquid using a shader in Unity.

Terrain is captured with a Kinect v2 depth sensor and the terrain game object in Unity is created from this depth array. Based on the depth a specific texture is applied — I made realistic, volcanic (with lava instead of water), colorful, colorful with contour lines and exotic designs. Another module is a calibration UI, where the user can calibrate the platform to correctly project onto the terrain.

Hand gesture recognition was done from depth data, because the projection overlays the hand which made it hard to recognize using a normal camera. I reached 60% recognition accuracy, which was unfortunately not enough for the final version.

Compiler

This project involves designing and developing a compiler for IFJ22. The compiler takes code written in IFJ22 (a language similar to PHP) and translates it into IFJcode22 (similar to Assembly), a low-level programming language.

The compiler consists of four parts. The lexical analyzer generates tokens based on a finite state machine I designed — tokens are sequences of characters (for, int, if, 5, ;, ...). These tokens are consumed by the syntax analyzer, which uses an LL grammar to verify the syntax and semantic correctness of the code (a top-down approach). For expressions (5 + 5 * 2) a bottom-up approach is applied: I designed a precedence analysis table which systematically applies the rules to the expression. Code is generated once the equivalent instructions from the input have been processed.

UDP and TCP client

This project was implemented to understand how packets move through a network. It improved my understanding of addressing and routing. I only implemented the clients and used a simple server for testing.

First I handled command line arguments, where the user can specify mode (TCP/UDP), host, port and IP. I created a socket based on the mode argument, then the necessary structs for addressing the host when sending and receiving packets. There is a main loop which, based on the type of communication (receive or send), handles and expects certain packets. In TCP a three-way handshake is executed first. At the end all structs in memory are freed.

In TCP each packet is acknowledged with an ACK packet from the receiver to ensure no packets are lost. UDP is used for video and audio streams, for example, where one missing packet won't have a big impact, so packets are not acknowledged. To verify the correctness of the packets, Wireshark was used.