Software Engineer (AI)

What matters most

• Production LLM automation

  The role designs, builds and operates tool-calling agents, orchestration layers and RAG. Prepare to explain an end-to-end system design, including how it moves from a model response to a usable workflow.

• Integrations and workflow design

  The posting calls for MCP-server and external-model integrations, alongside automation approaches such as n8n or Claude Code-style agents. Prepare to explain when an orchestrated workflow is preferable to a more autonomous agent.

• Backend and operational ownership

  The engineer builds scalable services and APIs for automated workflows and customer-facing features, and owns infrastructure, observability and performance across AI and non-AI services. Prepare an example that connects service design with production operation.

• Product and engineering foundations

  The role works with product, design and data to solve customer problems and helps define engineering practices and AI safety and reliability norms. Prepare to show how a technical decision could account for user experience and establish a reusable practice.

Likely challenges to think through

• Automation versus dependable outcomes

  Preparation hypothesis, not a stated employer problem: in a tax-focused product, more autonomous tool use may increase workflow coverage while making incorrect outputs harder to catch. Prepare to discuss where validation, evaluation and human review would fit, without assuming Numeral uses a particular control.

• Model flexibility versus operational complexity

  Preparation hypothesis: supporting external providers and MCP-style integrations can speed iteration but add failure modes and latency. Prepare to compare a flexible orchestration design with a simpler path, including how you would observe failures and decide when added complexity is justified.

• Fast delivery versus durable architecture

  Preparation hypothesis: defining foundations in an early-stage environment may require choosing between a quick customer-facing workflow and a more extensible service. Prepare to explain a decision framework using the posting’s stated trade-offs: latency, cost, reliability and iteration speed.

Topics to prepare

• Agent and integration architecture

  The role calls for tool-calling agents, model orchestration, MCP-style integrations and external model providers. Suggested revision exercise: diagram an agent workflow, marking tool permissions, provider boundaries and what happens when a tool fails.

• RAG and retrieval choices

  Production RAG is an explicit responsibility; vector database concepts such as Pinecone and HNSW are listed as a bonus. Suggested revision exercise: compare keyword and vector retrieval for a hypothetical tax-document question, including how you would check whether retrieved material supports the answer.

• Production tradeoffs and reliability

  The posting emphasizes latency, cost, reliability, iteration speed, observability and AI safety norms. Suggested revision exercise: talk through two designs for the same LLM workflow and explain what you would measure, where you would place human review, and when you would change designs. Human review is a suggested safeguard, not a stated requirement.

• Workflow and service boundaries

  Numeral wants scalable APIs and automated workflows, with n8n-style orchestration mentioned. Suggested revision exercise: sketch which steps belong in a backend service versus a workflow engine, and explain retry and failure boundaries.

• Product decisions in a tax setting

  The role involves customer-focused work with product, design and data teams; tax or regulatory experience is preferred, not required. Suggested revision exercise: frame a hypothetical customer tax-workflow problem, identify questions for those teams, and explain how you would prioritize speed against correctness without assuming any particular tax rule.

Skills to revise

• Build a tool-calling prototype

  Suggested exercise: implement a small agent that invokes a mock service through a defined tool interface, then handles a failed call and an invalid response. This practices the posting’s hands-on agent and integration work; use Cursor or Claude Code during development to revisit the AI-tool usage it requests.

• Implement a small retrieval flow

  Suggested exercise: assemble a few sample documents, retrieve relevant passages and have an LLM answer with references to those passages; test a question the documents cannot answer. This practices the explicitly named RAG responsibility without assuming Numeral’s data or implementation.

• Build an API with PostgreSQL and Redis

  Suggested exercise: create a small workflow API that persists task state in PostgreSQL and uses Redis for a clearly defined temporary state or cache; explain how it behaves on retries. Backend APIs, PostgreSQL and Redis are explicit in the posting.

• Instrument and assess a workflow

  Suggested exercise: add request tracing, latency and error measurements to the prototype, then inspect a simulated provider failure and estimate its cost impact. The role owns observability and performance and makes cost and reliability tradeoffs; these specific measurements are preparation suggestions.

• Connect a TypeScript interface

  Suggested exercise: build a small TypeScript interface that submits a workflow request, displays progress and presents a recoverable error. Frontend experience, especially TypeScript, and customer-facing features are stated in the posting.

• Map your evidence to the detailed requirements

  The role specifies 7+ years building products and 2+ years building with LLMs. Prepare a brief timeline of relevant work, plus concrete examples involving PostgreSQL, Redis, cloud services, TypeScript, and AI development tools such as Cursor or Claude Code.

• Prepare an end-to-end AI system walkthrough

  Choose a production example you can discuss. Diagram the user need, model orchestration, tool calls or retrieval, integrations, deployment, and what you personally built and operated.

• Practice a hypothetical tax-automation design

  As an interview exercise—not a claim about Numeral’s implementation—sketch an agentic workflow for a tax-related task. Show its API boundaries, tool or MCP-style integration points, human review, and how it handles uncertain outputs.

• Explain your architecture tradeoffs

  Prepare one decision record comparing latency, model cost, reliability, and iteration speed. Include how you would observe performance and diagnose failures across AI and non-AI services, which the posting says this role owns.

• Show how you evaluate safety and reliability

  Bring a compact test plan for an LLM workflow: representative cases, tool-call failures, retrieval errors, and criteria for escalation. Frame it as your proposed approach to the AI safety and reliability norms the role would help define.

• Rehearse a customer-focused collaboration story

  Prepare one example showing how you worked through ambiguity with product, design, or data partners, chose a scope, and used customer feedback to refine a shipped feature.

• Write questions about the role’s foundations

  Ask which AI workflows are in production, how the team measures reliability and customer value, where architectural ownership begins, and what would matter most for this hire to establish first.