AI Engineer – Trust & Explainability (AI Platform)

AI Engineer – Trust & Explainability (AI Platform)

Job Summary

The AI Platform team is building the shared runtime that every AI agent at MeridianLink will run on as part of the MeridianLink One platform: model gateway, orchestration, memory, and observability, delivered as reusable services so product teams build agents once and run them safely for hundreds of credit union and lender customers. This role builds the trust and explainability layer of that runtime: the tracing, evaluation, and explanation capabilities that let engineers understand exactly what a multi-agent workflow did and why, and that let product teams show end customers the right level of explanation to trust an agent's output.

This role has a strong dotted-line relationship to MeridianLink's Security Operations team, so the platform's trust guarantees and Security's threat models are built from the same picture.

About the Opportunity

This is a deeply technical, hands-on engineering role and an integral part of the platform itself. You will write the code that traces agent actions across a multi-agent workflow, adopt and extend leading open-source explainability and observability frameworks, and build the primitives product teams use to surface explanations to lenders, credit unions, and their borrowers. You should have real hands-on experience with LLM-based systems and a strong pull toward the problems of trust, tracing, and evaluation. You do not need to be an expert in all of it. This is a place to build that expertise while the platform is being built, working alongside a Staff engineer team lead and architects who will invest in your growth.

What it means to be an L2 AI Engineer at MeridianLink

L2 AI Engineers handle a broad range of work independently, from bug fixes to feature development. They have developed a specialization and are deepening it. They write code that other engineers trust, ask good questions, and are starting to help the engineers around them. L2 AI Engineers at MeridianLink are active, daily users of AI-assisted development tools.

Technical Execution & Delivery

Completes assigned features and bug fixes independently, with limited need for day-to-day guidance

Designs components within a well-defined scope; escalates questions on complex system design rather than guessing

Participates in code review and provides constructive feedback

Surfaces blockers proactively rather than waiting for check-ins

Craft & Professionalism

Writes tests that cover the functionality they ship

Monitors and responds to issues with their own work

Documents decisions and implementation details that others will need later

Agent Tracing & Explainability

Understands how to instrument an agent so a single conversation can be followed from the first user message through every model call, tool call, retrieval, agent handoff, and decision to the final response

Builds the correlation that ties actions across multiple agents in one workflow into a single readable trace

Turns raw trace data into an explanation a human can follow: what the agent did, what it relied on, and why it chose that path

Customer-Facing Trust & Explanation

Understands that the explanation an engineer needs and the explanation a borrower or loan officer needs are different, and builds for both

Works with product engineers on what an agent should disclose at the product surface: what it did, what data it used, how confident it was, and what a human should verify

Applies judgment about the right level of explanation for a regulated lending and account-opening context

Trust & Explainability Frameworks

Familiar with the open-source landscape for LLM observability, tracing, and evaluation, and can evaluate a framework against the platform's needs

Integrates and extends existing frameworks rather than rebuilding them, and builds what does not exist yet

Keeps the platform's instrumentation aligned to emerging standards so traces stay portable across tools

Agent Evaluation & Quality

Understands how to measure the quality of non-deterministic output: golden datasets, rubric and LLM-as-judge scoring, and regression baselines

Builds evaluation checks that run in CI so a model swap, prompt change, or tool change is tested before it reaches customers

Treats evaluation data as versioned, reviewed code

Safety & Tenant Isolation

Knows the common attack patterns against LLM applications (prompt injection, jailbreaks, tool misuse, data exfiltration) and how to write tests for them

Understands tenant isolation as a platform guarantee and builds tests that prove one customer's agent can never see another customer's data

Contributes to the shared guardrail layer so every agent inherits protection without re-implementing it

What Success Looks Like

In the first few months, a successful hire has shipped tracing that follows a full agent workflow end to end, readable by an engineer who did not write it, and has contributed working components to the platform's evaluation harness. Over the first year, success looks like this: a first version of the customer-facing explanation primitive is live in at least one product, evaluation and isolation checks run automatically on every agent release, and when an engineer asks why an agent did something, or a customer asks whether to trust it, the platform answers. Engineers who thrive here like the intersection of AI and rigor, want to become the team's go-to on agent tracing and explainability, and are motivated by making AI something people can see into and trust.

Key Responsibilities

Multi-Agent Tracing & Explainability

Build tracing across the platform's gateway, orchestration, memory, and tool layers, following defined designs from the team lead and architects

Build the correlation that links actions across agents in a single multi-agent workflow, including handoffs, parallel branches, and retries

Build the developer-facing trace view that makes a full agent conversation readable to someone who did not write the agent

Build the explanation layer that turns trace data into a human-readable account of why an agent did what it did

Customer-Facing Trust

Build the platform primitives product teams surface to end users: explanation records, confidence and provenance metadata, and "what the agent relied on" summaries

Partner with product engineers on the Document Request Agent and the MLM agents to land those primitives in real products

Iterate on the explanation format based on feedback from product teams and customer-facing staff

Trust & Explainability Frameworks

Evaluate and integrate open-source observability, tracing, and evaluation frameworks into the platform runtime

Extend those frameworks where our agent workloads need more than they offer, and contribute fixes and extensions back upstream where it makes sense

Build the gaps: the pieces of trust and explainability tooling the ecosystem does not provide yet

Evaluation Tooling

Build and maintain components of the platform's evaluation framework: golden dataset management, test runners, scoring pipelines, and regression reporting

Build the tooling that helps teams create and version golden datasets from de-identified real traffic and synthetic cases

Run model and prompt comparisons for the platform's shared components and report what changed

Safety & Isolation Testing

Build red-team and adversarial test suites covering prompt injection, jailbreak attempts, tool misuse, and data exfiltration, and run them in the platform's release process

Build the automated test suite that proves agents on the shared runtime cannot cross tenant boundaries through memory, retrieval, tool calls, or model context

Share red-team findings and new attack patterns with Security Operations and pull their threat models into the platform's tests

Collaboration & Growing Others

Participate in design discussions and code reviews; give and rec