ML/AI Engineer

We are looking for a ML/AI Engineer who combines hands-on production engineering with a genuine scientific understanding of modern deep learning. This is not an API-integration role. You will train, fine-tune, optimize and serve models yourself — and you will be expected to explain, from first principles, why the choices you make are the right ones.
The requirements below describe the expertise we consider essential. We have deliberately kept the list narrow: everything here matters.

Engineering Stack:
PyTorch — confident, day-to-day working proficiency, including custom training loops, distributed training and debugging at the tensor level.
Python asynchronous programming — deep practical command of asyncio, concurrency patterns, streaming, backpressure and the failure modes specific to async code.
FastAPI — designing and operating production-grade asynchronous inference and orchestration services.
Hugging Face ecosystem — transformers, peft, and Unsloth for efficient fine-tuning.
DevOps for GPU workloads — GPU cluster orchestration on Kubernetes, Docker, and strong command of Linux as a working environment (not just as a place where containers run).

Deep Learning Fundamentals
You must have a solid, explainable grasp of classical deep learning, including:
Modern convolutional architectures and layer design.
The attention mechanism and transformer architecture in depth.
Tokenization — algorithms, trade-offs, and their practical consequences.
Fundamentals of computer vision.
Encoder, decoder and encoder–decoder architectures, and the ability to articulate clearly where each is the appropriate choice.
Stochastic gradient descent, learning-rate schedules, the forward pass and backpropagation, and the family of modern optimizers.
Numerical precision and quantization — FP16, BF16, FP8, FP4 and related techniques — including what each costs you in accuracy and what it buys you in throughput and memory.

LLM Architectures and Training
We expect awareness that extends well beyond proprietary models and API providers. You should follow open-weight model releases and the current research literature, and be able to discuss:
The differences between Mixture-of-Experts (MoE) and dense LLM architectures, and their respective serving implications.
The core ideas behind RoPE (Rotary Position Embeddings), MLA (Multi-head Latent Attention), GQA (Grouped-Query Attention) and KV-caching.
Modern fine-tuning pipelines and their infrastructure, including FSDP (Fully Sharded Data Parallel).
Post-training and alignment methods, including GRPO (Group Relative Policy Optimization) and RLVR (Reinforcement Learning with Verifiable Rewards).

Agentic Systems and Inference
Familiarity with current research on AI agents, agent harnesses and tool-use patterns.
The ability to build asynchronous ReAct- and Thinking-style agents with tool calling entirely from scratch, without depending on a framework such as LangGraph or CrewAI. Frameworks are acceptable as a convenience; they are not acceptable as a substitute for understanding.
Practical understanding of inference performance: KV-caching and prefix reuse, time-to-first-token, and the behaviour of inference engines such as vLLM and Ollama.
Experience with LLM tracing, logging and observability in production.
Working knowledge of MCP (Model Context Protocol) and A2A (Agent-to-Agent) protocols.

A background in academic research, university-level lecturing, peer-reviewed publication, or public speaking and conference presentation in the field.

5000-8000 €/mon. gross

Gynėjų g.14, Vilnius Calculate your travel time to work