About the opportunity

We’re looking for a hands-on AI Architect to establish the technical foundation of a new agentic platform for accounting firms.

We’re building an agentic system that triages accounting documents arriving through multiple channels, including email, WhatsApp, SMS and a mobile app. It classifies invoices and receipts, extracts structured data, communicates with clients and routes uncertain cases to human reviewers.

You will own the AI architecture—from long-running agent orchestration and open-weight model deployment to evaluation, observability, guardrails and production reliability.

What you'll do:

• Own the long-running agentic pipeline, including orchestration, state, model routing, retries and fallbacks

• Build an evaluation layer that lets us compare and replace models without losing visibility into quality

• Deploy, test and optimize both managed and open-weight models in production

• Design deterministic guards and validated boundaries around probabilistic AI components

• Make the pipeline fully observable across quality, hallucinations, latency, cost and failures

• Develop RAG and retrieval workflows using customer-specific knowledge and vector databases

• Extend the human-in-the-loop review flow and ensure feedback improves the system safely

We're looking for someone who has:

• Production experience automating business workflows with AI, supported by measurable improvements.

• Deep experience with agent frameworks such as LangGraph, Google ADK or Claude Agent SDK.

• Experience building stateful, long-running agents involving tool use, agent loops and human intervention.

• Production model deployment through Bedrock, Azure AI Foundry, Vertex AI or a comparable platform.

• Hands-on experience deploying and evaluating open-weight models such as Llama.

• Strong experience with evaluators, LLM-as-judge, hallucination measurement and confidence calibration.

• Experience with AI observability, deterministic guardrails, RAG and vector databases.

• Experience building MCP servers or clients, rather than only consuming existing integrations.

• Familiarity with A2A and agent-interoperability patterns.

• Experience with vLLM, OpenRouter, Hugging Face, Langfuse, Arize Phoenix or pgvector.

• Advanced use or extension of agent harnesses such as Pi, Kimchi or OpenCode.

• Experience with multimodal document processing, OCR or intelligent classification.

• Experience with accounting, ERP, GDPR or other sensitive and regulated workflows.

What you'll get:

• The opportunity to build a product from the ground up

• Direct collaboration with experienced founders

• Significant influence over technical architecture and engineering decisions

• The chance to build real-world AI applications that solve meaningful business problems

• A fast-moving environment with minimal bureaucracy

• Office based in Belgrade, working closely with the founding product team.

Why this role is different:

This isn't a role where you'll be implementing predefined tickets. You'll work to explore problems, develop prototype solutions, and build a platform that combines AI, workflow automation, and integrations into a seamless experience for accounting firms.

If you enjoy building products from scratch, experimenting with AI, and creating software that fundamentally improves how businesses operate, we'd love to hear from you.