🌟 About You

You’re the person teams trust when the workload is messy, privacy-sensitive, and “it must not go down.” You like building systems that stay boring in production: hardened VMs, predictable deploys, clean rollbacks, useful alerts, and security controls that work by default.

You can talk comfortably about trade-offs, latency vs cost, isolation vs operational complexity, GPU vs CPU, shared vs per-tenant, and you’ve probably earned that comfort by running something real in production and being on the hook when it breaks.

🚀 About Omnilex

Omnilex is a fast-moving AI legal tech startup born out of ETH Zurich. Our interdisciplinary team of 14+ people builds an AI product that helps lawyers and in-house legal teams research faster and answer complex legal questions with confidence. We combine external/public sources, customer-internal knowledge, and our own AI-first legal commentaries to tackle real-world legal complexity, often under strict data residency and privacy expectations.

Tasks

🧭 The mission

Own the private AI runtime that powers our product: the infrastructure, security posture, and operational reliability of running LLM + agent workloads on Swiss-based VMs (optionally with GPUs). You’ll make it safe, observable, scalable, and easy for the team to ship changes without fear.

🛠️ What you’ll build and run

1. LLM + agent runtime in production (Swiss/Local VMs)

• Design and operate a VM-centric architecture for agent execution and model serving (single-node vs multi-node, concurrency, streaming).
• Manage model artifacts: downloads, storage, integrity checks, versioning, and safe rollback paths.
• Implement pragmatic controls for runaway agents (timeouts, token limits, tool permission boundaries, sandboxing patterns).
• Make capacity predictable: tokens/sec, queueing/backpressure, p95 latency targets, peak load behavior, and graceful degradation modes.

1. Linux / VM / network fundamentals (done right)

• Establish hardened VM baselines: users/sudo, SSH posture, patching approach, sensible defaults.
• Apply resource controls (cgroups/ulimits), disk + IO tuning, and repeatable “why is this VM slow?” investigation playbooks.
• Own TLS decisions end-to-end (termination, cert lifecycle, internal mTLS where it matters), plus egress controls and private networking.
• Build debugging muscle for real failure modes (DNS, certs, MTU, packet loss, noisy neighbors / CPU steal, memory pressure, IO wait).

1. Observability and incident readiness for LLM services

• Instrument what actually matters for LLM workloads: tokens/sec, queue depth, context length distribution, timeouts, error classes, saturation signals.
• Turn logs/metrics/traces into action: dashboards that get used, alerts that don’t spam, and runbooks that work at 3am.
• Drive incident hygiene: triage patterns, mitigation tools, postmortems that result in concrete fixes.

1. Security & privacy as engineering, not paperwork

• Build a real secrets lifecycle (creation → distribution → runtime access → rotation → revocation).
• Enforce least-privilege access on VM-centric infra (service identities, scoped credentials, audit trails).
• Prevent sensitive data leakage via prompts/logs/traces (redaction, sampling discipline, “never log this” guardrails).
• Reduce supply-chain risk: container/dependency hygiene, provenance checks for model artifacts, scanning and patch workflows.
• Prepare for uncomfortable scenarios (suspicious outbound traffic, prompt injection leading to attempted exfiltration) with detection + response playbooks.

1. Safe delivery and change management

• Implement deploy strategies appropriate for AI runtime infra: canaries, rollback triggers, maintenance windows, and fast revert paths.
• Keep systems from drifting: config drift detection, baseline enforcement, and reproducible deployments (IaC/automation).
• Support database/environment changes that intersect with runtime reliability (migrations, customer splits, environment promotion).

What success looks like (first months)

• A clear map of the biggest reliability + security risks in the runtime, with a prioritized plan and measurable improvements.
• Repeatable deployments with rollback confidence and a sane patching cadence (OS + drivers + runtime).
• Dashboards/alerts that catch real incidents early (and stop waking people up for noise).
• Practical privacy boundaries: everyone on the team knows what data can/cannot leave the VM, and the system enforces it.

Requirements

📌 What we’re looking for

✅ Minimum qualifications

• Hands-on experience running production infrastructure (VMs, networking, Linux) for a SaaS or platform product.
• Strong operational skill: debugging, incident response, log/metric-based investigation, and making recurring problems disappear.
• Solid security fundamentals applied pragmatically: secrets, least privilege, egress control, dependency hygiene, auditability.
• Comfort automating everything that should not require humans (provisioning, deploys, checks, drift detection, runbooks).
• Clear communication and an ownership mindset, you can partner with product/dev without becoming a blocker.
• Proficiency in English.
• Full-time availability; Zurich-based with at least two on-site days per week (hybrid).

🎯 Preferred qualifications

• Experience operating LLM serving/agent systems (or similarly spiky, latency-sensitive workloads).
• GPU operations familiarity (VRAM sizing intuition, quantization trade-offs, fallback modes).
• Azure experience (identity, networking, observability, IaC).
• Familiarity with our ecosystem: TypeScript, Node.js, NestJS, Next.js.
• Exposure to ISO 27001-style environments or supporting security audits.
• Swiss work permit or EU/EFTA citizenship.
• Working proficiency in German.

Benefits

🤝 Benefits

• High-impact ownership over the most sensitive part of the stack: private AI runtime + reliability + security.
• A sharp interdisciplinary team working at the intersection of AI and law.
• Autonomy: you define the guardrails that let everyone ship faster with fewer surprises.
• Compensation: CHF 8’000–12’000 per month + ESOP, depending on experience and skills.

If you want to own the infrastructure that keeps privacy guarantees real and production predictable, even under LLM chaos, apply by pressing the Apply button.