kacper leszczynski

I’m an Applied AI Engineer with 3 years of software engineering experience, focused on building AI-powered workflows and backend systems in Python and FastAPI. I’m especially drawn to evaluation, latency, traceability, deterministic control, and real-world failure analysis—because strong models must behave predictably in production. My work spans LLM/RAG orchestration, structured outputs, and multimodal audio pipelines.

At TCL Research Europe as an Audio AI Engineer, I architected an end-to-end automated movie dubbing workflow integrating ASR, forced alignment, LLM-based text processing, TTS, voice cloning, and audio/video rendering. This reduced turnaround time by 75% and improved perceived quality by 0.2 MOS. I also evaluated ASR, alignment, translation, and speech-generation failures across challenging real-world audio to identify recurring error patterns and quality bottlenecks.

I’ve built real-time speech translation for RayNeo X3 Pro smart glasses by combining VAD-based segmentation, chunked inference, and trained streaming S2TT models under strict latency and device constraints. I reproduced and adapted a neural source-separation model for real-time TV deployment, benchmarking TensorRT and quantized inference to reduce compute time by 16% while preserving model quality. I then built a real-time vocal boosting service that adapts dialogue enhancement from ambient microphone input while maintaining audio–video synchronization.

Previously at Santander, I delivered backend and frontend functionality for internal enterprise workflows using Java, Kotlin, Spring Boot, Angular, PostgreSQL, and REST APIs. I designed API/database/application changes and optimized complex queries and processing workflows, reducing execution time by up to 80% and improving responsiveness. Alongside my work, I’m pursuing an M.Sc. in Artificial Intelligence and Machine Learning at Warsaw University of Technology, and I enjoy turning prototypes into dependable systems.