Staff AI Engineer

About Meridian Softworks Meridian Softworks is building the AI platform for the tax and accounting profession. Our product, Halo, runs the full engagement lifecycle for CPA firms (engage, gather, prep, deliver) and layers an AI assistant, retrieval, and agentic automation across all of it. We work with some of the largest accounting firms in North America, in a domain where the output has to be right: a misclassified form or a missed deadline is not a bug, it is a client's tax return. We are building non-deterministic, agentic systems for professionals who cannot accept errors, at the volume and time pressure of tax season. Getting that right is a genuinely hard systems problem, and it is the problem this role owns. The Role As a Staff AI Engineer, you are a technical leader for the Halo platform. You set architectural direction for how we build agentic systems, define the patterns and guardrails the rest of the AI engineering team builds on, and are accountable for the reliability and trustworthiness of what we ship, not just the code you personally write. This is a hands-on staff role. You will still design and build the hardest parts of the system yourself. But your leverage comes from the decisions that shape everyone's work: the agent-orchestration and state model, the evaluation and observability harness, the multi-tenant isolation and data-integrity guarantees, and the build-versus-buy calls on core platform components. You will move fluidly across backend, data, infrastructure, and product, and you will raise the engineering bar of the people and vendor teams around you. You will partner directly with product, AI, and firm-facing teams, and your technical decisions will connect straight to customer outcomes and hard commercial deadlines. What You Will Own These are the areas where you will set the technical direction, define the standards others build against, and be accountable for the result. Agentic workflow orchestration. The architecture for multi-agent coordination across tax document workflows with human-in-the-loop oversight: agent state machines, tool routing, context windowing, and retry semantics for processes that run for minutes or hours. You define the orchestration patterns the team reuses, not one workflow at a time. Workflow state management. State hydration for long-running agentic workflows, failure handling, checkpoint and resume, and recovery across distributed services. You own the durable run-record and state layer that makes long-running agents auditable and resumable, and the contract every agent builds on top of. Document intelligence at scale. Production-grade pipelines that extract, classify, and validate tax forms and financial documents across dozens of formats and quality levels. You set the architecture for accuracy, coverage, and the validation layer that turns raw extraction into output a firm can trust. Evaluation and observability. The measurement backbone: task completion rates, accuracy attribution, cost tracking per action, and regression detection, with outcomes attributable to specific agent reasoning steps when something goes wrong. You stand up the eval and observability discipline as a platform capability, so that "is this good enough to ship" is a number, not an opinion. Trust and reliability. Making non-deterministic agent output trustworthy for professionals who cannot accept errors: supervision layers, validation rules, and human review gates, plus the provenance and audit trail that lets a firm defend its own work. This is a first-class architectural mandate, and you own the standard for it. Cost, accuracy, and latency. Optimizing the trade-offs across document types, complexity levels, and client tiers during peak tax-season volume, and setting the framework the team uses to reason about them rather than tuning case by case. What We Are Looking For You have built and shipped production systems, and then kept them running. You have taken features from requirements through implementation, testing, deployment, and monitoring, and you have handled the incidents that follow. You can point to enterprise-grade products that users rely on today, and you can architect solutions end-to-end when correctness and uptime matter. At Staff level, you do more than deliver your own work. You have set technical direction on complex distributed systems and been right often enough that teams follow your lead. You define patterns, standards, and interfaces that make other engineers faster and safer, and you have the judgment to make build-versus-buy calls on core infrastructure and defend them. You reduce key-person risk instead of becoming it: you mentor engineers, raise the bar in design and code review, and leave systems more legible than you found them. You understand concurrency, failure modes, data integrity, and why things break at scale, and you reason clearly about non-determinism where perfect answers are not available and the job is to bound and manage the uncertainty. The boundarie