| Languages | Python at expert level. TypeScript or JavaScript at working level. |
| ML stack | PyTorch, Hugging Face Transformers, TRL, PEFT, Accelerate, Datasets |
| Training | SFT, LoRA and QLoRA, distributed training, dataset curation |
| Alignment | Hands-on work in at least one of RLHF, DPO, GRPO, or RLVR, with measured results |
| Agents | Tool calling, MCP, agent loops, context management, agent evaluation |
| Inference | vLLM, TGI, or llama.cpp. Quantisation and latency tuning. |
| Backend | FastAPI, async Python, Pydantic, PostgreSQL, Redis |
| Data | Embeddings, vector databases, RAG pipelines, chunking strategy |
| Operations | Docker, Git, Linux, CI/CD, GPU basics (CUDA, VRAM planning) |
| Evaluation | Benchmarks, custom eval sets, LLM as judge, regression tests |