DevOps / Platform Engineer - Scale AI & High-Volume Infrastructure

Smalk AI; Paris, Bordeaux; Hybrid; EUR 60,000 - 75,000; Full time

Benefits offered by this job

Permanent CDI

Paris (Station F) / Bordeaux

Hybrid or full remote

Direct CTO access

Ownership from day one

Job summary

Smalk AI is seeking a senior infrastructure engineer to scale a production platform handling billions of events. You’ll own throughput, latency, and capacity planning across analytics, ad serving, and AI workloads, using Kubernetes on bare metal, Terraform and Ansible for IaC, and a modern tech stack including PostgreSQL and a ClickHouse‑like analytics DB.

Remote options exist with a Paris base. You’ll lead partitioning, sharding, and security work, partnering with the CTO to ensure SOC 2

Qualifications

• Experience scaling high‑throughput systems to order‑of‑magnitude growth.
• Production Kubernetes experience and bare‑metal deployments.
• Tuning PostgreSQL under high load and managing data at scale.
• Familiar with Kafka, Redis and real‑time analytics pipelines.
• Awareness of SOC 2 or ISO audits and secure scaling practices.

Responsibilities

• Own throughput and latency of analytics and ad‑serving paths as volume grows.
• Design and implement data partitioning, sharding and retention strategies.
• Lead AI workloads pipelines: queues, GPU inference, vector search, cost per unit.
• Manage the platform: Kubernetes, IaC, capacity planning, autoscaling, upgrades.
• Deliver CI/CD with zero‑downtime deployments and production‑mirror environments.
• Improve observability: metrics, logs, alerting, runbooks, incident reviews.
• Drive security and compliance toward SOC 2 compliance.

Skills

High-throughput systems

Kubernetes in production

PostgreSQL tuning

Kafka

Redis

SOC 2 awareness

Team leadership

Tools

K3s on bare metal

Terraform

Ansible

Prometheus

Grafana

Loki

GitHub Actions

ClickHouse-like analytics DB

GPU inference

Job description

Most infrastructure jobs at startups mean inheriting a fire. This one doesn’t. Our platform runs in production today and does its job: more than 2.5 billion events and requests a month, over 70 million a day, around 1 ms p50 and around 10 ms p99 on the hot path. By the time you join, our own Kubernetes platform on bare metal will be live. The architecture is designed, documented and measured.

What we don’t have yet is the person whose whole job is scale. We expect analytics volume to grow 10x within six months, 100x within a year and 1,000x within eighteen months. That is the job: making sure the platform is ready before the traffic is.

What already runs

• Analytics at volume: 2.1 billion tracked events a month across our publishers, the largest French news sites among them. AI agent visits, other bot traffic and human page views, collected through our tracker and publisher CDN logs.
• Real-time ad serving: our ad serving API answer some 80 million requests a month.
• AI workloads every day: LLM answer analysis, embeddings, rerankers on GPU, ad generation and agents that work for our customers.
• A platform we chose and designed: Kubernetes (k3s) on bare metal, databases on dedicated machines next to it, observability, autoscaling per workload, written runbooks.

The future we’re building

Smalk is building the economy of that new world. We help publishers thrive in the AI Search era with a brand-new revenue stream, and help brands earn their place at the top of AI answers. That’s the category we’re pioneering: Generative Engine Advertising (GEA). Every publisher we sign brings its traffic onto the platform, which is why the volume curve looks the way it does.

Your first 12 months

Months 0 to 3: make it yours. Learn the platform by owning it: finish moving the last services onto Kubernetes, sharpen our integration environment (copy-on-write clones of PG database, deployments per branch), switch off the legacy PaaS. By the end, nothing runs that you couldn’t rebuild from the repository.

Months 3 to 6: the first 10x. Analytics goes from 2 billion to over 20 billion events a month. You find the next bottleneck before it finds us: ingestion, Kafka, the analytics database, Redis, autoscaling, and what each of them costs per billion events.

Months 6 to 12: design for 100x and 1,000x. Database size becomes the interesting problem: partitioning, retention, sharding, and the read and write paths that go with them. In parallel, you lead security and compliance, with SOC 2 as the target.

Why this is a good infrastructure job

• You start from a working system, not a rescue. Your time goes to scaling, not firefighting.
• The growth targets are written down. You know what scale you are building for, and when.
• Real volume from day one: billions of events, real-time serving and AI workloads, not a staging cluster waiting for users.
• Bare metal and Kubernetes you control end to end, chosen for performance per euro, not a black box you rent.
• Hard problems with a clear owner: sharding, data at 1,000x, cost per event, SOC 2. That owner is you.
• A measurement culture: we benchmark before we decide, and every decision in our docs comes with the numbers behind it.

What you’ll own

• Throughput and latency of the analytics and ad-serving paths as volume grows 10x, then 100x.
• Data at scale: PostgreSQL and a ClickHouse-like analytics database (around 500 GB today), tuning, replication, tested restores, then partitioning and sharding.
• AI workloads: queues running LLM calls and agents, GPU inference, vector search, and their cost per unit of work.
• The platform: Kubernetes, infrastructure as code, capacity planning, autoscaling, upgrades.
• Delivery: CI/CD, zero-downtime deploys, migrations and rollbacks, an integration environment that mirrors production.
• Observability and on-call: metrics, logs, alerting, runbooks and incident reviews.
• Security and compliance, all the way to SOC 2.

You’ll report directly to our CTO. Every architecture decision so far is documented with its reasoning, and every one of them is yours to challenge.

Python, Django, Celery and FastAPI on the application side. Kubernetes (k3s) on bare metal, Helm, Ansible, Terraform, PostgreSQL, a ClickHouse-like analytics database, Kafka, Redis, Prometheus, Grafana and Loki, GitHub Actions. GPU inference and LLM providers through APIs.

How we work

We believe developers are AI Agent Team Leads, not craftsmen typing code line by line. That holds for infrastructure too: you’ll work with Claude Code and our in-house agents to write manifests, read logs, review changes and draft runbooks. What we expect from you is judgement: knowing what will break at the next order of magnitude, measuring before deciding, and leaving systems that others can operate.

You’ll probably fit if you’ve already

• Taken a high-throughput system through an order of magnitude of growth, and know which part broke first.
• Run Kubernetes in production, ideally on machines you were also responsible for.
• Operated PostgreSQL under real load: tuning, replication, a restore you had to perform, a connection limit you hit.
• Partitioned or sharded a database that had outgrown a single machine, or can explain how you would.
• Migrated something live without losing data, with the rollback ready.
• Kept infrastructure reproducible as code, and kept a cloud or GPU bill under control by measuring it first.

You’ll need fluent French and working English.

Engineering school, university, self-taught: we care about what you’ve run in production, not where you studied.

Bonus points if you’ve run Kafka at scale, ClickHouse or another columnar or time-series database, GPU or LLM inference workloads, or Celery under heavy load, worked through a SOC 2 or ISO 27001 audit, or operated bare metal at a European provider.

What we offer

• Permanent contract (CDI).
• €60k to €75k (gross) per year depending on profile.
• Paris (Station F), Bordeaux, hybrid or full remote, with regular team touchpoints.
• Direct access to the CTO and full ownership of the platform from day one.
• A working platform with three orders of magnitude of growth ahead of it.
• Pioneering a new category and joining an AI-native company in the front row of AI Search.

Hiring process

We’ll ask you to walk us through a system you scaled and what broke on the way, then to tackle a hands‑on case: how you would take our analytics pipeline from 2 billion to 200 billion events a month, and what you would change first.