Designing and building simple application architectures in Python (API, data layer, processing), LLM-based tools (local models, RAG, embeddings)

Selected Projects:

Lexory – LLM-powered grammar learning backend with RAG
· Architected a multi-container Docker Compose environment orchestrating FastAPI (Swagger UI), Qdrant vector store, LanguageTool, and Ollama, with environment-aware URL rewriting for Docker vs local development.
· Designed hybrid retrieval pipeline with 2-step deduplication (category check + semantic similarity) to optimize storage and reduce redundancy; uses Qdrant client and local LLM for private grammar processing.
· Configured CPU-only PyTorch via dependency configuration to keep the project runnable on systems without GPU.

Chat on Gradio – Local LLM chat interface with model-aware UX
· Designed layered architecture to separate UI, orchestration, and inference logic for maintainability and extensibility.
· Designed modal-free inline confirmation flow for destructive actions, and implemented UI-state filtering pipeline to exclude non-conversational messages from model context, reducing prompt pollution and mitigating hallucinated responses.
· Developed license-aware interaction flow adapting UI behavior to model requirements (e.g., gated models like Gemma require explicit terms confirmation).
· Built IP-based multi-session persistence system with JSON storage, supporting session lifecycle management and safe deletion with confirmation.

MP3VG – PySide6 desktop app for generating two-language audio
· Converts paired items into structured MP3 using gTTS, FFmpeg, and pydub.
· Implemented configurable playback sequencing and pause control.
· Decoupled language sequencing via ACTIVE_PROFILE-driven configuration, enabling safe language switching without modifying core logic.