Alicja Golisowicz

● Developed novel benchmark addressing critical gaps in LLM evaluation by manually designing cognitive reasoning tasks focused on object tracking and spatial reasoning, demonstrating how models that excel at complex STEM problems struggle with fundamental cognitive skills easily manageable by humans ● Analyzed performance degradation patterns, identifying critical limitations in current LLM architectures with accuracy drops from 89% to 31% in extended contexts ● Conducted systematic evaluation of five small instruction-tuned LLMs (DeepSeek-R1-Distill-Qwen 1.5B, Llama-3.2-1B/3B, Phi-3.5-Mini, and Qwen2.5-1.5B) using multiple prompting strategies

● Led the development of a TTS system (English & Polish), improving fluency and naturalness across voices and accents. ● Maintained client-facing TTS infrastructure using ONNX and Triton; served models on GPU servers and balanced legacy support with model upgrades ● Upgraded the TTS pipeline with newer models (NaturalSpeech, YourTTS) and initiated an internal voice cloning prototype for multi-speaker synthesis. ● Conducted R&D for a character-based game project involving voice cloning and custom speaker adaptation. ● Engineered system prompts to guide GPT-3.5 and GPT-4 output according to CEFR English proficiency levels, enhancing a learner-facing chatbot. ● Built and integrated APIs for chatbot applications using LLMs and dynamic prompt chaining. ● Created and annotated a dataset for argument and objection classification to support client-specific NLP applications. ● Developed extractive summarization pipelines for internal research and real-world call transcript processing in customer service scenarios. ● Delivered internal and client-facing NLP/TTS demos using FastAPI and Streamlit.