- Architect agentic systems — plan, memory, tool use, multi-agent delegation, evaluation loops, guardrails. Pick the right abstraction for the problem, not the one on the hype curve.
- Push model capability into production. Design prompt and context strategies, tool interfaces, retrieval and reranking, structured output, streaming, and evaluation — across text and multimodal inputs (vision, documents, audio).
- Own the evaluation story. Build offline eval sets, online LLM-as-judge loops, regression harnesses. Know the difference between a metric that moves your users and a metric that moves only your dashboard.
- Squeeze the system. Prompt caching, batching, speculative decoding, model routing, token budget management, latency targets. Know your P50/P99 and why they look the way they do.
- Contribute upstream. Read SDK source when docs are thin, open PRs against open-source agent frameworks, write crisp bug reports when a vendor's orchestration service returns a weird 500.
- Mentor and set the bar. Your design reviews, code reviews and technical writing shape how the rest of the team thinks about agents.
- Bachelor’s degree in Artificial Intelligence, Computer Science, or a related field.
- 0-3 years of experience as an AI/AI Agent Engineer.
- Solid ML fundamentals. You can explain transformers — attention, positional encoding, KV cache, tokenisation, sampling — without hand-waving. You've read at least a few core papers in full, not just the abstracts.
- Deep LLM application experience. Multiple production systems built on top of frontier models (Anthropic, OpenAI, Gemini, open-weight). You know the practical edge cases: tool-use stability, structured-output failure modes, long-context degradation, prompt-injection defence, cost control.
- Agent systems depth. You've built something with real agent behaviour — planning, memory, tool orchestration, multi-step execution, error recovery — not a single prompt in a loop. Experience with multi-agent coordination (delegation, sub-agent protocols, MCP-style tool servers) is a strong plus.
- Multimodal experience. Hands-on work with vision-language models, document AI (OCR, layout, tables), or audio — end-to-end from ingest to grounded output.
- Strong engineering craft. Python at a senior level — async, typing, testing, packaging, observability. Able to read and navigate a large codebase. Git hygiene that makes reviewers' lives easier.
- Production AI engineering. You can take ML, LLM, embedding, vision, or multimodal models from prototype to production — designing APIs and inference services, building RAG/embedding pipelines, containerizing workloads, handling CPU/GPU deployment, and optimizing latency, throughput, reliability, and cost.
- AI platform & MLOps experience. Hands-on experience operating AI workloads in production with model/version management, CI/CD, evaluation gates, observability, autoscaling, rollback, and failure handling. Experience with Docker, Kubernetes/AKS, Azure AI services, GPU inference, or model-serving frameworks such as vLLM or NVIDIA Triton is a strong plus.
- Fluent with modern coding agents. You use Claude Code / Cursor / Copilot / equivalents daily as a force multiplier. You understand where they shine and where they fail, and you can design prompts, context and tool boundaries to get the most out of them.
- Communication. Writes and speaks clearly in English.
This position will be contracted through Bosch’s external vendor under a 1-year contract. Salary and benefits will be discussed during interview