Own the evaluation layer for our agents — harnesses, metrics, golden datasets, regression gates, and human-in-the-loop review.
Architect and implement agent workflows using LangChain, LangGraph, or equivalent; state memory, routing, tools registries and recovery paths.
Own systems from experimentation through production and operate in production: tracing, monitoring, latency, cost-per-task budgeting, on-call for what you build.
Diagnose real production failures: hallucination, tool misuse, retrieval misses, silent degradation and turn each into a durable fix and a test.
Set the technical standard for agent work at Hilbert: review designs, define patterns others build on, and raise the bar across the team.
Collaborate closely with the founding team and cross-functional partners — communicating tradeoffs, progress, and technical decisions with clarity.
Make pragmatic engineering decisions under ambiguity—ship, learn, iterate.
Our Current Hurdles
These are the kinds of problems you'll walk into on day one:
Intelligent retrieval across heterogeneous approaches — our agents need the right information at exactly the right moment. The challenge isn't picking one retrieval method; it's combining RAG, graph-based retrieval, and other approaches into a unified strategy that fetches the most relevant content precisely when the agent needs it — no more, no less.
Agentic workflows that solve real-world problems — it's building workflows robust enough to handle the unexpected. When an agent hits an edge case, missing data, or a situation it wasn't explicitly designed for, it needs to reason through it — leveraging available context, escalating to a human when it can't, and never silently failing.
Evaluation beyond vibes — we need systematic, reproducible evals that actually predict real-world performance. If you've built custom evaluators for RAG or agent workflows, we want to talk.
