Trách nhiệm chính

Own the LLM integration layer used for document extraction (OCR), accounting‑entry proposals, and the in‑app chat agent, currently built on Anthropic/OpenAI APIs.
Implement prompt caching to cut redundant token usage – an identified ~83% cost‑reduction opportunity – and introduce per‑tenant usage budgets and quotas.
Harden the agent against prompt injection: treat all extracted document text as untrusted input, and require explicit confirmation gates before any tool call that grants data access, creates a contact, or moves money.
Replace the agent’s currently hard‑coded (effectively disabled) permission model with real, enforced per‑tool authorization checks.
Evaluate and implement model tiering – cheaper models for extraction, stronger models for judgment calls – and plan a fallback provider for resilience.
Build an evaluation / golden‑file test harness so model or prompt changes can be validated before shipping to production.
Implement PII redaction before data is sent to third‑party LLM providers, and partner with the Full‑Stack Engineer to log every agent tool call to an audit trail.
Advise on data‑processing‑agreement and data‑residency questions for the AI sub‑processors (Anthropic, OpenAI), alongside the company’s data‑protection advisor.
Contribute to the n8n‑side automation where it touches AI extraction – the document‑intake pipeline – coordinating handoffs with the Full‑Stack Engineer.

Yêu cầu công việc

3+ years of experience building production applications on top of LLM APIs (Anthropic Claude, OpenAI, or equivalent).
Hands‑on experience with agentic tool‑use patterns – function/tool calling, multi‑step agent loops – not just single‑shot prompting.
Strong prompt‑engineering skills, including context management and techniques for defending against prompt injection.
Proficiency in Python or TypeScript, sufficient to build and modify a backend AI‑integration layer.
Experience building or using evaluation frameworks / golden‑file testing for LLM‑based systems.
Practical understanding of LLM cost structures and optimisation techniques – prompt caching, context trimming, model tiering.
Working knowledge of security concerns specific to LLM systems: prompt injection, data exfiltration via generated content, unsafe tool authorization.
Clear written communication in English; able to work asynchronously across time zones.

Yêu cầu ưu tiên

Experience with OCR / document‑extraction pipelines specifically – receipts, invoices, financial documents.
Familiarity with Swiss accounting concepts (MWST/VAT codes, KMU chart of accounts), or willingness to learn them quickly.
Experience with n8n or a similar workflow‑automation tool.
Background in a regulated or compliance‑sensitive domain (fintech, healthtech, legal tech).

Hình thức làm việc

Thông tin khác

Remote AI Agent Engineer