AI/agents engineer

What is Nanzen

Nanzen is disrupting B2B by bringing agentic workflows to finance operations, automating critical, trust-sensitive flows that still run on emails, spreadsheets, old databases, manual approval chains and basic wire transfers.

We are ambitious, we are building Nanzen to be an era-defining company, where there's a clear before and after in B2B commerce.

We prefer in person but we do hybrid. Life can be better if you can make the best of both worlds. The days were not sitting next to each other we miss the speed of a whiteboarding session and the good vibes.

The Nanzen team

Nanzen's philosophy is to bring the smallest team of the greatest people that cover maximum footprint.

You need to have the ambition to work on something small that can change the world. This requires your attention and love. It's about caring, not sucking your life into non-existence. We have families and we intend our kids to see their parents but this is not a 9 to 5 job at Dunder-Mifflin.

Trust is everything. At Nanzen we base our relationship in transparency, autonomy and accountability (= allergic to BS).

The AI/agents engineer role

What you own

Agent design. Trajectories, tool surfaces, and the decomposition of a finance workflow into steps a model can do reliably. Making numerical reasoning verifiable rather than merely plausible.

The trust harness. Graders, confidence scoring, reflection and escalation logic, error taxonomies.

Context. An agent is only as good as the state it decides from. Reconciling ERP, CRM, contracts and email into one coherent representation, keeping the financially material facts as it grows, and retrieving the right ones into a bounded window.

Evals and observability. Regression suites that catch drift before a customer does. Unmeasured means unshipped.

Inference and performance. Cost and latency as product constraints, not tuning done later. Model choice per role, budgets per turn, and the honest tradeoffs between them.

Customer reality. Sitting with finance teams to find where the genuine ambiguity lives, and moving them up the autonomy ladder.

The bar. You set the standard for whoever joins next, and you help choose them. We hire slowly and stay small on purpose.

You'll probably recognise yourself

• You've shipped LLM systems people depend on, not demos.
• You care more about measurement than about the demo.
• You're comfortable when correctness is a financial requirement, not a quality goal.
• You'd rather delete an agent than ship one you can't explain.
• You've been burned by a benchmark that went green while the product got worse, and you have opinions about why.

Probably not for you if

• You want to work on models rather than on systems built out of them.
• You'd rather add a capability than prove the last one holds.
• You need the problem fully specified before you start. Most of what we know, we learned by shipping it to a finance team and watching.

How we work

Very AI-pilled but anchored in reality. Small team, high trust, no ceremony.

Stack: Python, FastAPI, Postgres, Next.js/TypeScript, multi-provider LLM routing, OpenTelemetry tracing, Terraform, GCP/AWS