AI Systems Engineer

ABOUT VAILENT

Vailent is the AI infrastructure for the materials industry — chemicals, polymers, elastomers, rubber. The companies

in this space run on a mess of CRMs, ERPs, point tools, and flat files. We're replacing all of that with one system that

turns every interaction, transaction, and physical asset into usable commercial data.

Materials are the foundation of the physical economy: they're in everything. Every product humans build, ship, eat,

wear, or drive starts here. But the industry is still massively under-instrumented, running on fragmented tools and the

institutional knowledge of people who've been doing it for decades. At Vailent, we're building the infrastructure that

will transform this industry for the next century, capturing multi-modal industry context across both software and

hardware.

About the Role

A full-stack platform engineer who can run a multi-app B2B platform end to end — by directing fleets of AI agents and

verifying everything in the real environment. You'll own the whole stack: cloud infrastructure, backend, frontend, data,

and deep enterprise-ERP integration. The job isn't writing code with AI; it's operating it — decompose, fan out, verify

adversarially, ship.

One seat doing what's normally three or four.

We run a B2B platform spanning roughly ten applications on a shared cloud backbone, with deep integration into

customers' enterprise systems (SAP/ERP). This role owns it end to end — from the Terraform and IAM underneath to

the React components on top, and the SAP RFC calls in between.

The differentiator isn't typing speed. It's the ability to hold an entire platform in your head and conduct AI agents

through it without dropping correctness — shipping across many repositories at once while keeping the architecture

coherent. AI orchestration here is not a productivity add-on; it's the core multiplier that makes the scope possible. We

hire for that fluency, and for the discipline that makes it safe.

What You'll Do

• Own the platform end to end. Multiple applications plus shared SDKs on a single cloud backbone —
• React/TypeScript front ends, FastAPI/Python services, the Terraform/IAM/ECS infrastructure underneath, and a
• shared design system.
• Stand up infrastructure and environments from scratch. New services, cloud accounts, tenants, connectors,
• data syncs, migrations (including cross-region) — provisioned and proven, never just stood up and assumed.
• Direct fleets of coding agents. Decompose a cross-repo change into disjoint tasks, fan them out to parallel
• agents in isolated worktrees, run adversarial multi-reviewer passes, then reconcile the results.
• Integrate with enterprise systems at depth. SAP/ERP integration via RFC/BAPI — reading and where
• necessary authoring ABAP, reverse-engineering business rules, handling sales-order and customer-master flows,
• currency/unit/sales-area mapping, and idempotent event sync.
• Architect multi-tenant data. Postgres row-level security as the tenant-isolation core, JSONB-backed
• tenant-extensible capability platforms (custom fields, validation, masking), careful migrations, and a graph
• database where it fits.
• Ship at volume without losing coherence. Multiple PRs across multiple repos in a working session, CI green,
• deployed and verified — while keeping the design clean.
• Author the thinking, not just the code. Specs, design docs, discovery-question sets, and runbooks that let work
• be understood and resumed by others.
• Build the tooling that makes AI effective here. Per-codebase navigation maps, documentation indexes, guard
• hooks, and custom skills — invest in making agents good at this codebase, then reap it on every task after.
• Automate yourself forward. Treat every repeated task as a bug to be fixed. When a workflow recurs, capture it as
• a reusable Claude skill, hook, or slash command so the next run — yours or a teammate's — is one step instead
• of ten.
• Review like an adversary, deploy like a surgeon. Catch the regression the happy path missed, separate “it
• renders” from “the data is correct,” refute false blockers, and touch shared state only with a reason and a green
• light.

How We Work

Hire for the disposition. The stack is learnable; this isn't.

These principles are non-negotiable, because at this volume they're what keep the work correct. If you don't already

work this way, the throughput becomes a liability instead of an asset.

01 — Prove it in the real environment. “Done” means demonstrated, not asserted. A green badge over $0 /

insufficient data is a failure. subrc=0 means nothing until the record reads back. The data wins, never the badge.

02 — Never guess. Verify what's knowable in the code; ask about what's a genuine product decision; assume

nothing in between. Confident fiction is worse than an honest “I don't know yet.”

03 — Diagnose before you touch. “Look into it” means read-only until told to fix — especially on anything live. Root

cause and a proposed fix come first; the change waits for an explicit go. Production is sacred.

04 — Copy what works. If working examples already solve a problem, read the proven pattern and adapt it. Don't

invent a fresh approach and burn an afternoon proving it wrong.

05 — Enhance in place, never fork. Generalize the existing path — add an optional parameter where today is the

degenerate case — rather than shipping a parallel reimplementation. Design the capability; a single customer is the

validating example, not the spec.

06 — Risk isn't size. Bigger isn't worse; riskier is. Risk is load-bearing code modified × silent-failure potential × blast

radius. A large additive change can be safer than a one-line edit to a hot path.

07 — Build to scale — or name the debt. Ship the agreed slice now, but flag anything that won't scale as explicit,

revisit-able debt. Hardcoded shortcuts are fine only when chosen out loud, never smuggled in.

08 — Own the correction. Verify findings adversarially — a second pass whose job is to refute the first. When the

evidence turns, reverse yourself out loud. The best catches are corrections of your own confident conclusions.

09 — Words are a feature. Terminology has precise internal meaning. Inventing loose language for things that

already have names is a real defect — caught and corrected on the spot, not waved through.

10 — Leave a trail. Every session ends with a handoff so the next one — human or agent — starts informed. Specs,

runbooks, tracked tickets, and durable notes are part of the deliverable, not overhead.

The Environment

Frontend — React, TypeScript, Vite, TanStack Query, vitest, a token-based design system, Playwright for

verification.

Backend — Python, FastAPI (async), SQLAlchemy, Alembic, Celery, Pydantic; an SNS®SQS event bus with

idempotent dedup.

Data — PostgreSQL with row-level security, schema-per-app, JSONB + GIN/GIST, Neo4j (Cypher), pgvector.

Platform / Infra — AWS (ECS Fargate, Aurora, RDS Proxy, Route53, ACM, WAF, CloudFront, IAM/OIDC),

Terraform, dual-account, per-branch Docker stacks, gitflow.

Enterprise integration — SAP ECC via RFC/BAPI, ABAP, pyrfc, customer/order master data, additional ERP

connectors, M2M auth.

Identity & AI — Auth0 (Organizations, M2M, custom claims), JWT entitlement gating; Claude Code agents,

worktrees, skills, hooks, MCP.

Must have

• Fluent AI orchestration. You already run agents in parallel, isolate their work in worktrees, and verify their output
• adversarially — not “I’ve used Copilot.”
• Genuine full-stack + infra range. Comfortable going from a React component to a Postgres RLS policy to a
• Terraform module in the same day.
• Systems debugging instinct. You chase root cause across service boundaries — auth, pagination, dependency
• conflicts, integration mismatches — and don't stop at the first plausible story.
• The evidence reflex. You distrust green badges, demand real fixtures, and prove things with a working
• screenshot, a read-back record, or a live payload.
• Self-correction. You can describe a time you reversed your own confident conclusion because the evidence said
• so.
• An automation reflex. You instinctively turn recurring work into reusable Claude skills, hooks, and commands —
• raising your own efficiency floor instead of re-doing toil.
• Operating discipline. Read-only until authorized, copy proven patterns, enhance-in-place, precise language,
• clean handoffs.
• Thick skin & plain speech. You take blunt, fast feedback well and explain your reasoning simply.

Nice to have

• Enterprise ERP / SAP depth. RFC/BAPI, ABAP, customer & order master data — or the nerve to
• reverse-engineer a customer's system to that depth.
• Tooling-builder streak. You've built the scaffolding that makes other agents and engineers effective: nav maps,
• indexes, skills, guard hooks, templates.
• Architectural taste under constraint. You reach for the boundary that keeps future cost flat, and can name why a
• rewrite or a scatter is the wrong move.
• Multi-tenant / B2B context. Tenancy isolation, per-tenant configuration, and the failure modes they bring.
• Compliance fluency. GDPR / SOC 2 / ISO 27001 — comfortable with ROPA, control mappings, and runbooks.
• Design-system literacy. Tokens over hardcoded values; able to run a UX and a UI pass on your own work.

#vailent