Head of AI Agents
Lima Xi's intelligence layer is moving from models to agents. This role owns that transition — designing the autonomous systems that monitor corridors, trace cascades and put a recommended action in front of an operator.
From models to agents
Lima Xi turns geopolitical and capacity signals into corridor-level forecasts. Today much of that pipeline is orchestrated. The direction of travel is agentic: systems that decide what to investigate, pull the sources they need, reason about a cascade, and surface a recommendation with its own confidence attached.
That is a genuinely hard engineering problem in a domain where a wrong answer delivered confidently is worse than no answer at all. An operator hedging fuel or rerouting pharma freight acts on what we tell them. Reliability, traceability and calibrated uncertainty are not nice-to-haves here.
You will own the agent architecture end to end — design, evaluation, and the production systems that keep it honest.
Company leadership
The mandate
01Agent architecture
- Design the agent systems that monitor corridors, investigate signals and trace cascades
- Own orchestration, tool use, retrieval and memory across long-running workflows
- Decide where an agent is the right abstraction and where deterministic code is better
- Build for traceability — every recommendation must be explainable back to its sources
02Evaluation & reliability
- Build the evaluation harness: how do we know an agent got it right, at scale?
- Design confidence calibration so the platform reports uncertainty honestly
- Establish regression testing against historical corridor events and backtests
- Own failure modes — hallucination, tool misuse, silent degradation
03Production engineering
- Ship agent systems into production with monitoring, cost controls and latency budgets
- Work with the data platform team on retrieval quality and source freshness
- Optimise for cost per decision, not just capability
- Keep the stack current as the underlying model landscape shifts
04Product partnership
- Translate operator workflows into agent behaviours that fit how people actually work
- Partner with the CTO and VP Engineering on roadmap and technical direction
- Bring the outside view — what is genuinely state of the art versus what is hype
What we're looking for
Essential
- Strong software engineering fundamentals — you ship production systems, not notebooks
- Hands-on experience building LLM-based agents: orchestration, tool use, retrieval
- Demonstrable rigour on evaluation — you have built eval harnesses, not just vibes-tested
- Python proficiency and comfort across a modern data and inference stack
- Clear written English — you document design decisions others can follow
Strongly preferred
- Experience with agent frameworks and their limitations, and opinions about both
- Retrieval-augmented generation at production quality, not demo quality
- Working with structured and time-series data alongside unstructured sources
- Cost and latency optimisation for inference-heavy workloads
- Prior work in a domain where being wrong has real consequences
Character
- Sceptical by default — you test claims, including your own
- Comfortable saying an approach will not work before it burns a quarter
- High ownership across the full lifecycle: design, ship, monitor, fix
- Pragmatic about the model landscape — no attachment to a particular vendor or framework
A genuine plus
- Exposure to logistics, aviation, trading or another time-sensitive decision domain
- Experience with forecasting or causal inference alongside language models
- Open source contributions in the agent or evaluation space
- Mandarin or Thai
Own the agent layer, not a corner of it
This is not a role optimising prompts inside someone else's architecture. You are defining how Lima Xi's intelligence layer works, at the point where the company is committing to agents as the direction.
- Genuine architectural ownership from day one
- A real domain problem — corridor cascades are hard, and the ground truth is measurable
- Equity participation in a seed-stage company with category ambition
- Work directly with the founding team, including a CTO with three prior exits
- Remote-friendly with Hong Kong as home base
- Freedom to choose the stack, provided you can defend the choice
Before you apply
Is this a research role or an engineering role?
Engineering, with research literacy. We need production systems that run reliably and cheaply, informed by someone who genuinely follows the field. If you want to publish papers, this is not the right role.
What does the current stack look like?
Python-based, with the intelligence pipeline built around structured corridor and capacity data alongside unstructured geopolitical sources. Specifics are discussed under NDA.
How much freedom is there on architecture?
Substantial. You will own the agent design. We care that you can defend the decisions, not that you pick a particular framework.
Why does evaluation get so much emphasis here?
Because operators act on our output — hedging fuel, rerouting freight, committing capacity. A confidently wrong recommendation has a cost. Calibrated uncertainty is a product requirement.
Send it to careers@limaxi.io
Applications handled confidentially. Technical conversations are with the CTO and VP Engineering directly.