AI Engineer

Location
Valencia
Working hours
Full-time
Work mode
Remote
Level
Employee
Department
Other
Posted

Apply on Lãberit (in Spanish)

We are looking for a Senior Full-Stack Engineer dedicated to developing our MarIA platform: a web application that lets healthcare professionals configure, monitor and manage our real-time voice-to-voice AI agent.

You will work with our core team, making sure the application scales seamlessly to tens of thousands of simultaneous AI conversations.

Responsibilities:

  • Design and maintain the harness of our clinical agents: agentic loop, tool registry, token budget management, stop conditions, context compaction and retry policy.
  • Define the right multi-agent topology for each problem (subagents with isolated context, skills, handoffs, router) and justify why that one and not another.
  • Build long-horizon planning loops with state persisted to the filesystem: plan files, resume checkpoints and human checkpoints before irreversible actions on patient data.
  • Package clinical procedural knowledge as Skills (SKILL.md, progressive disclosure) and external capabilities as our own MCP servers on top of HIS/EHR, with permissions by risk level (read / write / destructive).
  • Set up and grow our agent evaluation layer: regression datasets, LLM-as-a-judge with auditable criteria, automated evals that block changes to the harness and security evals of the harness itself (indirect injection via clinical content, timeouts, tool overuse), using LangSmith, Braintrust, Langfuse, promptfoo, DeepEval, Inspect or similar.
  • Securely integrate our systems with hospitals and electronic health record (EHR) platforms, including the permissions model and traceability of what the agent reads and writes.
  • Instrument the agents with per-span traces (Langsmith, OpenTelemetry, Sentry, or similar): cost, latency per tool, success rate per step, transcription quality and prompt cache hit rate.
  • Develop and evolve the web application through which healthcare professionals interact with our agents, with the judgment to direct and review generated code.
  • Improve the backend infrastructure and automate deployments with Docker and GitHub Actions, taking care of the developer experience and the isolated environments where the agents run.
  • Take part in key architecture decisions: what is solved with a model, what with a deterministic harness and what with plain code.

What we are looking for:

  • 3+ years developing production software, with at least 1 year taking LLM-based systems beyond the demo: not chat wrappers, but systems with tools, state and real failures.
  • Proven experience designing tool interfaces for agents: typed schemas, input validation, idempotency and errors the model can use to correct itself.
  • Deep understanding of context management: what goes into the window and why, compaction, external memory in the filesystem, cache-aware ordering.
  • Knowledge of MCP at the implementation level, not the user level: having written a server, understanding tools/resources/prompts, transports and authorization.
  • An evaluation-first rather than intuition-first mindset: experience defining metrics and evaluation datasets for non-deterministic systems with LangSmith, Braintrust, Langfuse, Ragas, DeepEval, OpenAI Evals or similar. If you can't measure whether a change improved something, it isn't a change.
  • Mastery of Python and TypeScript, and the ability to read code you didn't write —your own or generated— with judgment, in order to review and reject it.
  • Experience designing relational schemas and authentication and authorization flows (IE: Supabase).
  • Security judgment in agentic systems: indirect injection surface, least privilege, execution isolation, handling of sensitive data.
  • Solid knowledge of Docker and CI pipelines.
  • Fluent Spanish and English (spoken and written).
  • Live in Valencia or Madrid

Excellent communication and a strong sense of responsibility and autonomy.

It's a plus if you have:

  • Experience with applied fine-tuning (LoRA/QLoRA, SFT, distillation) and, above all, evidence of having evaluated whether it was worth it compared with adjusting the harness.
  • Experience serving your own inference: vLLM, SGLang, TensorRT-LLM or similar; continuous batching, quantization, GPU sizing and cost per token compared with an API.
  • Experience with ASR and voice models applied to the clinical domain.
  • Knowledge of real-time communications (WebSockets, WebRTC) and streaming APIs.
  • Public contributions: MCP servers, skills, agent tools, papers or benchmarks.
  • Knowledge of the healthcare sector and regulations such as GDPR, MDR or HIPAA.

Skills

  • High proactivity Initiative Collaborative Communication skills Strong organizational skills
Apply on Lãberit (in Spanish)