Slavisa S.

My elevator description of what I do best: I help implement generative AI models in projects to speed up workflows and processes and make them more efficient. I've worked extensively with chatbots (rag, multi step reasoning, escalation, etc.) and agents (using AI to perform complex and sometimes very long tasks). I specialize in ML architecture & design and can also craft a model from scratch to fit a very particular use case with strict requirements. I work across the full spectrum of modern AI, from training and fine-tuning models on my own hardware and cloud GPUs to shipping production systems on frontier APIs. Model training & research: I train and fine-tune LLMs and voice models locally and on RunPod, working in PyTorch with hands-on experience in distillation (MSE/KL/CE curriculum training), embedding-space architectures, and context compression research. I understand what's happening under the hood, not just at the API layer. Frontier AI systems: Deep, current working knowledge of the entire frontier stack — OpenAI, Anthropic/Claude, Gemini, Groq, Cerebras — including how to choose between them, benchmark them, and build provider failover so your product isn't married to one vendor. Voice AI: Production phone and browser agents at sub-300ms latency using Twilio, Deepgram, ElevenLabs, and Hume, with clean interruption handling and AI-to-AI testing pipelines. RAG & knowledge systems: Pinecone-backed retrieval, vector store Q&A, multi-language support, 10x+ query speed improvements over prior stacks. The short version: I can fine-tune a model on a GPU cluster, and I can also ship the product it powers. Most freelancers do one or the other.