Staff Software Engineer - Agent Architecture

Company Description

At PayNearMe, we’re on a mission to make paying and getting paid as simple as possible. We build innovative technology that transforms the way businesses and their customers experience payments. Our industry-leading platform, PayXM™, is the first of its kind—designed to manage the entire payment experience from start to finish. Every click, swipe or tap is seamless, fast and secure, helping non-commerce businesses boost customer satisfaction, accelerate payments, and reduce costs.

Our single platform handles it all: cards, ACH, digital wallets such as PayPal, Venmo, Cash App Pay, Apple Pay and Google Pay, and even cash at more than 62,000 retail locations nationwide. Today, thousands of businesses across consumer lending, iGaming and online sports betting, property management, and tolling trust PayNearMe to deliver a payment experience that drives real results.

In September 2025, we raised a $50 million Series E funding round to accelerate our growth.

We’re a team of 300+ employees across 41 states, headquartered in Silicon Valley with satellite offices in Dallas, TX and Holmdel, NJ.

Join us and be part of a team that’s shaping the future of payments—one experience at a time.

Responsibilities

We build agentic AI products that our customers interact with across different modalities. These agents sit on top of the same money-movement platform that handles real funds for businesses in regulated industries, so they have to be safe, compliant, and predictable in ways most consumer AI products are not.

We're looking for a Staff Engineer to own the architecture and implementation of these agents end-to-end. This is a builder role for someone who has shipped agents to production at scale, not just used them. You will define how we build agents at PayNearMe: the frameworks, the integration patterns with our existing systems, the guardrails around money and PII, and the testing/eval/observability loop that lets us improve agents safely over time.

Our core platform stack is Ruby on Rails with MySQL (monolith) plus Go microservices on AWS/Kubernetes, with Datadog for observability. Our agents interact with consumers and business partners across a wide variety of use cases.

What you'll do

You'll define and build the agent platform used across applications at PayNearMe—the architectural patterns, the shared infrastructure—and you'll set the bar for how agents are designed, tested, evaluated, and operated in a regulated, money-movement context.

• Own the architectural direction for agentic AI at PayNearMe in partnership with other engineering leaders. We are building an agent platform, not a single agent—our business customers have different rules, brand voices, allowed actions, knowledge bases, and compliance postures, and the architecture has to treat per-tenant configuration, isolation, and evaluation as first-class concerns.
• Produce and maintain architecture documentation (current state, target state, migration plan) and drive alignment across product, engineering, security, and compliance.
• Design, build, and ship production agents—including voice and chat agents for a wide range of payment-related activities—that integrate cleanly with our Ruby on Rails / MySQL platform and partner services (ElevenLabs, Twilio, and others).
• Treat tool design as a first-class discipline: tool schemas, descriptions, idempotency, side-effect semantics, and error surfaces directly determine agent quality, and for money-moving tools they determine whether we can stand behind every action the agent took.
• Make and defend the "what kind of intelligence goes where" decisions: when to lean on a partner's stack vs. orchestrate frontier LLMs directly, when RAG is the right answer vs. tool calls vs. fine-tuning, when a small/fast/cheap model is sufficient vs. when a frontier model is warranted, and where classical ML or deterministic logic is a better fit than an LLM at all.
• Design the seams that let us swap providers, voice vendors, and models as the landscape shifts—without rewriting the agents that sit on top of them.
• Design and operate the agent lifecycle as a closed loop: testing, offline evals, online evals, observability, scoring, and a disciplined feedback path from what we observe in production back into the test suite and eval set.
• Own the rollout discipline for non-deterministic systems: prompt and agent versioning, shadow mode, canary-by-tenant, gradual ramps, and rollback playbooks that account for the fact that the "bad version" may have already taken real payments. The system has to get measurably better over time, not just ship.
• Own the unit economics of agent interactions. Token budgets, prompt and semantic caching, model cascades (cheap model first, escalate on uncertainty), batch APIs, latency-vs-cost tradeoffs, and per-tenant cost attribution should be instrumented and reasoned about explicitly—at scale, the gap between a well-engineered conversation and a naive one is the business.
• Build the guardrails that make agents safe in a payments context: scope enforcement, refusal behaviors, deterministic handoffs for anything money-changing, PCI-compliant handling of card data, PII protection, and clear human-in-the-loop or fallback paths when confidence is low.
• Own the identity and consent model for agent-initiated actions—who the agent is acting as, when step-up authentication is required before a consequential action, and how explicit consent is captured and stored in a form that holds up in a dispute or chargeback.
• Treat prompt injection and social-engineering of the agent as real attack surfaces; stand up a red-team practice that exercises them continuously, especially against money-moving tools.
• Treat voice as its own modality, not a text agent with a microphone—design for latency budgets, barge-in and turn-taking, STT/TTS error modes, DTMF fallback, recording and consent, and the operational realities of telephony.
• Partner with Security, Compliance, and Legal to ensure agent behavior meets PCI-DSS, state-level payments regulations, and our customers' own compliance obligations. Make agent decisions reconstructable: for any consumer interaction we should be able to explain to a regulator, an auditor, or a disputing party exactly why the agent did what it did, on what information, and with what authorization.
• Raise the bar across the org for agent engineering: define shared patterns for prompts, tools, evals, telemetry, and incident response; serve as a reviewer and approver for architecture decision records (ADRs) and major designs in the agent domain.
• Teach the rest of engineering how to build, evaluate, and operate agents—most engineers on the team are picking this discipline up for the first time, and the team's velocity depends on how well that knowledge transfers.
• Partner with the Engineering Managers, Product, and other Staff peers to shape the roadmap—develop deep expertise in both the technical system and the business need (what our customers and their consumers actually want from an agent), and translate that into durable platform capabilitie