Job Description
AI Engineer – Agent Orchestration, Memory & MCP· Full-time · Hybrid / Remote
Most agent systems are stateless between sessions, or bolt on memory as an afterthought, memory is the product
We build the governed context layer for the agentic enterprise, turning complex, heterogeneous enterprise data into context graphs that power LLM reasoning pipelines for clients in pharma, finance, and defence. Our graph is designed to get smarter over time: every agent interaction, confirmed inference, and discovered rule feeds back as permanent knowledge
How that happens reliably and faithfully is an open design challenge, and the most interesting engineering problem on the team.
What you'll own
As Engineer #2 on the AI side, you'll have architectural ownership of four areas from day
one:Memory write-back, Design the mechanism that turns agent sessions into durable graph knowledge. Faithfulness, conflict safety, and scale all need solving. This is genuinely open territory, here and at most companies building in this space.
Multi-agent orchestration, Own the router, specialist sub-agents, streaming traces, and memory tier handoffs. Build for real production failure modes, not the happy path.
MCP integrations, Each enterprise system gets its own MCP server. Extend the connector library and own the gateway to client LLM stacks.
Agent guardrails, Access control enforced at the data layer, LTN formal compliance logic, and provable constraints — not prompt-level suggestions.
What we're looking for
You've shipped multi-agent systems in production (router patterns, real failure modes, instrumentation, not demos). You've thought hard about how knowledge gets captured, verified, and promoted, and what goes wrong. You've built MCP servers with auth scoping and dynamic tool discovery. You know when to use multi-hop graph traversal versus vector search because you've built both.
Production Python is second nature: type hints, structured logging, async FASTAPI.
Bonus
Experience with faithfulness eval / LLM-as-judge, entity resolution, bitemporal data modelling, write-back conflict resolution, multi-tenant namespace design, or formal/neuro-symbolic constraints (LTN). Regulated-industry data experience is a
Strong Plus
The write-back problem, making an agent system that learns trustworthily at scale — is unsolved here and almost everywhere else. You'll own the architecture across formal logic, bitemporal graphs, and production LLM pipelines, at a company where that's the core product, not a side project.