Samuel K.

From PyTorch to the front-end. I can work with and without AI. Both are essential, most people can only do one. My professional background is in low-level programming and systems work. I am comfortable with actual deep learning internals: PyTorch, JAX, transformers, autograd, diffusion models, computer vision, CUDA-adjacent optimisation. I have built transformers and an autograd engine from scratch, and made open-source tooling that ended up being used by a lot of AI researchers and engineers. Things I can help with: - LLM systems, agent workflows, tool use, RAG, embeddings, vector search - transcription pipelines, speaker diarisation, structured extraction from messy real-world audio - image generation systems, multiple model providers, inpainting, generation pipelines - Stable Diffusion inference, fine-tuning, LoRA work, ComfyUI-related tooling - computer vision models, including custom classifier training for production - distributed GPU inference, queues, storage, streaming updates, cost control - AI products where the hard part is not just the model, but making the whole contraption behave sensibly I also do the surrounding software stack. Usually this means Python, SvelteKit, TypeScript, FastAPI, PostgreSQL, Drizzle, Supabase, AWS, S3, and Vercel AI SDK. I am quite happy moving between model, backend, and frontend.