Sobre esta vaga de Senior AI Engineer, Agents na Nacre Capital
About EveryWatch
EveryWatch is the largest and most trusted data source in the secondary watch market. Established by a group of watch lovers, EveryWatch was created in response to the increasing popularity of luxury timepieces, with the aim of bringing unprecedented transparency and insight to the watch market. The first platform of its kind, EveryWatch combines all aspects of the watch world under one roof: a one-stop shop for watch collectors, vendors, and enthusiasts.
Position Overview
We're looking for Senior AI Engineer, Agents - an Agent Architect who will be responsible for designing, building, and scaling AI-powered solutions across EveryWatch. Working closely with our engineering, product, sales, and data teams, they will identify opportunities where AI and LLMs can automate complex processes, improve decision-making, and create new capabilities for our users and internal teams.
This is a highly hands-on role for someone who combines strong Python and LLM engineering skills with a creative, proactive mindset. You will not simply be given a roadmap. You will be expected to understand how the business operates, identify where AI can make a meaningful difference, propose new solutions, and take them from idea to production.
Our existing AI work, including WatchChat, provides a foundation to build on. The opportunity now is to expand that foundation into a broader ecosystem of intelligent agents that can work with EveryWatch's unique watch-market data and support collectors, dealers, sales teams, data operations, and engineering.
In short, we're looking for someone who doesn't just ask, “What can we build?” but “What should we build?”
Requirements
What you'll actually do
- Invent the agent roadmap. Sit with sales, data and product, find the repetitive expert work, and come back with a ranked list of agents worth building - with a real view on feasibility, cost and impact. This is the core of the job, and it doesn't stop after the first quarter.
- Build them yourself. You are hands-on. You design the architecture and you write the code - tools, loops, sub-agents, memory, state, evaluation. Not a spec-writer with a team underneath.
- Own the platform under the agents. Every new agent should be cheaper to build than the last: shared tool layer over EverWatch data (pricing, auctions, listings, references, portfolios), an MCP surface over our existing backend, shared memory, tracing, and a reusable eval harness.
- Harden WatchChat alongside us. Multi-turn state and memory, latency, cost, tool-call reliability, regression gates. It's live-bound and it has to stay right.
- Make quality measurable. Golden multi-turn datasets, programmatic verifiers for tool/argument correctness, LLM-as-judge on held-out sets. If we can't measure an agent, we don't ship it.
- Treat cost and latency as design constraints. Cheap models for routing and intent, strong models where they earn their keep; context budgets, caching, and - where it pays off - fine-tuning (SFT/LoRA on curated production trajectories) instead of ever-larger prompts.
What we're looking for
Three things, and we won't trade any of them away:
- Creative. You generate agent ideas the business hadn't thought of, and you can tell the difference between one that will work and one that demos well. You start from a use case, not a framework.
- Strong architect. You can design an agent platform that's still standing in five years - state, memory, tool boundaries, sub-agent decomposition, evaluation, failure modes, cost. You'll be asked to critique our current architecture in the interview, and we expect you to find things.
- Heavily hands-on. You ship. Deep production experience, writing the code yourself, at pace.
Must have
- 6+ years shipping production software, of which 2+ on LLM systems that real users touched.
- Real agentic depth: ReAct or equivalent loops, tool/function calling, planners, state & checkpointing, long-term memory, HITL steps, streaming. Not "I called the OpenAI API."
- Hands-on LangGraph (or a strong argument for something better) plus a tracing/observability stack - LangSmith, Langfuse or similar.
- Python in production: FastAPI, async job processing, clean service boundaries.
- RAG done properly: retrieval + re-ranking + relevance judgement, and honest evaluation of all three.
- Evaluation as a habit, not an afterthought (RAGAS/DeepEval/GEVAL, LLM-as-judge, pass@k).
- Cloud production experience - AWS (Bedrock, SQS, EC2/EKS, S3) or equivalent - with Docker and CI/CD.
- The spine pushes back on us. We explicitly want someone who asks "why did you build it like that?"
Nice to have
- LLM post-training: SFT with LoRA/QLoRA, preference/RL methods, reward and verifier design.
- Voice agents (LiveKit/Pipecat or similar), or vision-language work.
- Multi-agent orchestration and MCP integrations.
- Text-to-SQL over a real, messy production schema.
- Report/document generation agents - structured, sourced, client-ready output.
- Guardrails, PII handling, hallucination detection, risk scoring.
- Interest in watches, collectibles or market data. Not required - curiosity about the domain is.
Benefits
What you get
- Ownership of EverWatch's entire agent layer, and the roadmap for it — not a corner of someone else's.
- A dataset that doesn't exist anywhere else, and users who care whether the answer is right.
- Direct line to the CTO; decisions in days, not quarters.
- Budget for models, tooling, and the coding agents you want to work with.
- Competitive compensation, remote-first.
