Staff Engineer

Causa Prima; München; On-site; EUR 120,000 - 190,000; Full time

Job summary

Prima Mente seeks a Staff Engineer to own the architecture and operation of autonomous agents that handle accounts payable and receivable end-to-end. You will design, implement, and operate the system that validates invoices, resolves disputes, and optimizes cash flow across the platform.

You will lead architecture decisions, ensure zero-trust security, and drive data infrastructure improvements while collaborating with product and platform teams.

Qualifications

• 8+ years shipping production systems and architecture leadership.
• Strong systems design thinking across complex domains.
• Expertise in Python and TypeScript in production environments.
• Deep PostgreSQL knowledge: event sourcing, RLS, indexing, schema evolution.
• Hands-on with LLM-powered systems and agent orchestration.
• Experience with ERP/financial integration and data pipelines.

Responsibilities

• Design the multi-agent architecture and enforce the boundary between reasoning and enforcement.
• Lead agent dispute resolution and dynamic discounting workflows.
• Implement matching, reconciliation, and cash orchestration across payables/receivables.
• Define secure integration with ERPs, banking APIs, and document AI tools.
• Establish security, zero-trust, and compliance patterns across platforms.
• Contribute as a senior IC shaping platform and data infrastructure.

Skills

Systems design

Python

TypeScript

PostgreSQL

LLM-powered systems

LangGraph/LangChain

Security

AI tooling

Tools

LangGraph

LangChain

LlamaIndex

LlamaParse

Job description

Staff Engineer

You'll be the most senior engineer on the agents that run finance operations for our customers - designing, building, and operating the autonomous agents that handle accounts payable and receivable end-to-end. From invoice validation and dispute resolution to cross-company negotiation and cash optimization.

This is an individual-contributor role at the top of our ladder: no direct reports, full technical authority over the architecture the rest of the team builds on.

What you'll do

• Agent architecture - Design the multi-agent system: agent boundaries, event contracts, orchestration patterns, and the separation between LLM reasoning and deterministic enforcement. LLMs inform; rules enforce. You own where that line sits for every action that moves money, and the authorization gateway that makes it non-negotiable.

• Agentic dispute resolution - An agent that reads the claim, gathers the evidence from the PO, contract, and delivery record, forms a position, and works it to settlement - negotiating with the counterparty where that's appropriate, escalating to a human where it isn't. Deductions and short pays classified, tied to the underlying claim, and driven to resolution rather than parked in an aging bucket.

• Agentic dynamic discounting - Agents that continuously price early payment against the cash position on both sides of the transaction, propose terms, and execute them. Static payment terms become a negotiation that runs on its own.

• Matching and reconciliation - Three-way matching across purchase order, goods receipt, and invoice with configurable tolerances; cash application against bank statements and remittance advices, including partial payments and residuals; the exception paths that follow when any of it fails. Unglamorous but important: The foundation everything agentic sits on.

• Validation & anomaly detection - Business rule checks, duplicate and near-duplicate detection, pricing discrepancy detection, and anomaly detection across time. Precision matters as much as recall here: a false positive is a blocked supplier and an angry procurement lead.

• Integration framework - ERP systems first and deepest, then banking APIs, accounting platforms, Gmail, Drive, and Slack. Auth model, credential lifecycle, data normalization. Today each connector is a project; you make the next one a configuration.

• Compliance & security - Agents operate within a zero-trust model. You'll design how agents verify independently against source data, how outbound content is reviewed, and how the system prevents cascading failures from prompt injection. Plus the European e-invoicing mandates now landing across France, Germany, and Spain, and what EN 16931 means for the pipeline.

• Platform & data infrastructure - Own the primitives every agent action rests on: an append-only PostgreSQL event store with tamper-evident hash chaining for replay and audit; cryptographic signature verification on agent events and approval chains before anything external fires; the Workflow Constraint Engine that gates payments, emails, and API calls on deterministic policy, not model output; row-level multi-tenancy and tenant-scoped credentials across every data store; attribute-based authorization via Cerbos, evaluated the same way whether the caller is an agent or a human; and the MCP security layer - auth, capability-scoping, rate limiting - for external agent-to-agent access.

• The engineering standard - Architecture decision records, review bar, on-call, and a clear view of what we deliberately don't build. You'll write the things the next ten engineers read in week one, and you'll spend time with controllers and AP leads watching agents fail in ways no test predicted.

What we're looking for

• 8+ years shipping production systems; currently Staff, Principal, Lead IC, or a founding engineer. You've carried an architecture through a scaling step and lived with your own decisions.

• Strong systems design thinking - agent boundaries, failure modes, trust boundaries, contract evolution. Money movement is not retry-safe by default and you already know why.

• Depth in Python and/or TypeScript in production, and deep PostgreSQL: event sourcing, row-level security, indexing tradeoffs, schema evolution under load.

• Hands-on experience with LLM-powered systems, agent orchestration frameworks (LangGraph, LangChain, or similar), and document AI (LlamaIndex, LlamaParse, or similar).

• Strong opinions about when to use an LLM vs. a rules engine vs. traditional ML - and the courage to defend them.

• Experience designing data pipelines, integration architectures, or event-driven systems in production.

• Security awareness - you think about what happens when untrusted input meets your pipeline, and about segregation of duties before someone makes you.

• Technical leadership without a title: reviews that teach, designs that get adopted, engineers who get faster because you're there.

• Effective use of AI coding tools with strong review skills for LLM integration and auth code.

• Domain knowledge across the purchase-to-pay and order-to-cash cycle. You don't need all three areas below, but you need real depth in one and working fluency in the rest - enough that when a controller describes an exception, you understand it before they finish:

  • Accounts payable - PO and non-PO invoice intake, coding to GL account and cost centre, approval routing and delegation of authority, three-way matching and tolerance handling, blocked-invoice and exception workflows, duplicate detection, payment runs, payment terms and early-payment discounts, supplier master data and bank-detail verification, period-end accruals for un-invoiced receipts.

  • Accounts receivable - invoice issuance and e-invoicing, cash application and remittance matching, partial payments, short pays and residuals, credit memos, deductions and dispute handling, dunning and collections strategy, credit limits and credit risk, aging and DSO.

  • Procurement - requisition to purchase order, supplier onboarding and qualification, catalogues and contract pricing, goods receipt, off-contract and maverick spend, sourcing and supplier negotiation, and the handoff into AP where most of the pain actually lives.

• Fluency with the systems this runs on: ERPs (NetSuite, Dynamics, or the SMB stack), the sub-ledger to general-ledger relationship, and how period close constrains what an agent is allowed to do and when.

Nice to have:

• Peppol and EN 16931 in production, or bank connectivity via ISO 20022.

• Cryptographic primitives in production: digital signatures, hash chaining, tamper-evident logs.

• SOC 2, GDPR, or data-residency engineering in a regulated environment.

• You've founded something, or have been employee 1 to 5 in a startup.