We are looking for an LLM / Agentic Evaluation Rig Engineer to build the system that decides whether our AI output is good enough to ship. Because our commentary sits next to externally reported financials, we cannot rely on vibes — grounding, faithfulness, and hallucination have to be measured, tracked, and gated before anything reaches a customer.
You own the evaluation infrastructure: the datasets, the scorers, the harnesses, and the CI gates that hold the AI and agentic layers to a hard quality bar. You are the team's source of truth on whether a model, prompt, or agent change is actually an improvement — and the one who blocks it if it isn't.
What makes this role different
- You define "good enough to ship" — your gates block regressions in grounding and faithfulness from reaching production.
- Evidence over vibes — every claim is checked against the verified source data it must be grounded in.
- Agentic evaluation — you evaluate multi-step reasoning flows, not just single prompts.
- Real leverage — your rig is how the whole AI team moves fast without breaking trust.
Responsibilities
Datasets & Scorers (35%)
- Build and curate evaluation datasets, including adversarial and edge-case sets with ground-truth labels
- Build scorers for grounding, faithfulness, hallucination, factual consistency, and structured-output validity
- Combine rule-based checks, reference-based metrics, and LLM-as-judge where appropriate
- Verify generated claims map to verified source data — no unsupported statements
Harnesses & CI Gates (30%)
- Build harnesses that run evaluations reproducibly across model, prompt, and agent versions
- Wire evaluation into CI so grounding / faithfulness regressions block releases
- Track quality over time with dashboards and clear pass / fail thresholds
Agentic Evaluation (25%)
- Evaluate multi-step / agentic flows — routing, tool-use, verification, confirmation
- Build trace capture and step-level scoring for agent runs
- Detect where a flow silently degrades
Collaboration (10%)
- Partner with the Staff AI Engineer to turn findings into model / prompt / orchestration improvements
- Partner with QA to integrate AI evaluation into the broader release process
Technical Stack
Evaluation
- LLM eval frameworks (promptfoo, DeepEval, Ragas, LangSmith)
- LLM-as-judge, reference-based metrics
- Dataset / ground-truth curation
AI & Orchestration
- LLM APIs & managed LLMs (Bedrock / Vertex / Azure OpenAI)
- RAG & agentic patterns (LangGraph)
- Structured-output validation
Engineering
- Python
- CI/CD (GitHub Actions)
- Dashboards & metrics tracking
What You'll Build in Year One
- A labeled evaluation dataset suite (including adversarial cases) for the generation and agentic layers.
- A scorer library for grounding, faithfulness, hallucination, and structured-output validity.
- A reproducible harness wired into CI that blocks releases on quality regressions.
- Step-level trace capture and scoring for agentic flows, with dashboards leadership can trust.
Required Qualifications
Core
- 4+ years in software / ML engineering, with hands-on work building LLM evaluation or quality tooling.
- Real understanding of grounding, faithfulness, and hallucination — and how to measure them rigorously.
Technical
- Strong Python and solid engineering practices (reproducibility, CI/CD).
- Comfort designing evaluation for non-deterministic systems without producing flaky or meaningless metrics.
- Familiarity with LLM eval frameworks and LLM-as-judge patterns.
Nice-to-Have
- Experience evaluating agentic / multi-step LLM systems.
- Familiarity with RAG, structured output, and managed LLMs in-VPC.
- FinTech / financial-services domain or other high-stakes, correctness-critical AI.
- Background in statistics or measurement / metrics design.
