技能/経験

■Experience
・Mandatory
1. 5+ years building and running software in production (Python; AWS or equivalent cloud).
2. 2+ years hands-on shipping LLM-based or agentic applications to real users, not only proofs of concept.
3. Evaluation-driven practice: can show an evaluation set, error analysis and the iteration it drove on a real project.
4. Daily use of agentic coding tools (Claude Code, Codex or equivalent) as a core part of engineering work, including reviewing and testing their output.
5. Experience leading a technical workstream or mentoring engineers (formal team-lead experience not required; this is a hands-on role, ~80% coding / 20% coordination, suited to a strong developer ready to grow into a lead).

・Nice to have
Trained, evaluated or maintained supervised or deep-learning models in production.
Experience in pharma or another regulated industry .
Experience co-developing with external delivery partners or vendors.
Front-end development (React/Next.js) for agent user interfaces.

■License
・Mandatory
None. Bachelor's degree in Computer Science or a related quantitative discipline, or equivalent experience.

・Nice to have
AWS certification (Solutions Architect Associate or Developer Associate, or higher).
Master's or PhD in Computer Science or a related quantitative discipline.
DeepLearning.AI or equivalent LLM/agentic AI coursework.

■Skill-set
・Mandatory
Agent architecture: workflow vs. agent harness, tool calling, MCP, memory and context management, guardrails against prompt injection and data exfiltration.
At least one agent framework ( Bedrock AgentCore, LangGraph, OpenAI Agents SDK, or equivalent).
Grounding: RAG with vector stores and retrieval over structured data via SQL.
LLM foundations: model selection, tokenisation and cost, context window and caching, when to fine-tune or self-host.
Production operations for AI systems: observability and tracing, cost/latency optimisation, statistical regression testing in CI/CD, prompt/model versioning, infrastructure-as-code.
Software engineering fundamentals: full-stack application design, data modelling, testing strategy, secure-by-design practice (dependency scanning, secrets hygiene).
AWS in production: Lambda, ECS/Fargate, API Gateway, Bedrock, IAM.
Clear written and verbal explanation of technical trade-offs and uncertainty to non-specialists.

・Nice to have
Knowledge graphs or semantic layers over structured data (e.g. Snowflake).
Document-processing pipelines for Japanese-language material.
Customising agent environments: skills, MCP servers, hooks, shared context files (CLAUDE.md/AGENTS.md).
Eval/tracing tooling such as Langfuse, Braintrust or LangSmith.
Azure OpenAI and multi-provider gateway experience.