Senior SRE/ML Ops Engineer - Performance

About the role

The Performance team owns advertiser outcomes at Vibe: the prediction models, the bidding engine, pacing, and the delivery stack underneath them. You'd report to Clément Schiano, Engineering Manager for Performance.

We're hiring because our infrastructure work has become two jobs. One is building the platform and tooling that gets data scientists from experiment to production quickly. The other is keeping the production serving path fast, cheap and reliable as traffic grows. One person currently does both, and the reliability half loses whenever there's a modeling deadline. This role takes over that half.

Why it's interesting:

• Scale is the problem we have. Vibe was acquired by Walmart and the plan is to run many more advertisers through the same bidding stack. How much of that we can take on depends directly on what the infrastructure can absorb.
• Every bid is a real-time decision with advertiser budget behind it. If p99 moves the wrong way, advertisers under-deliver and we lose money that day. Cost works the same way: replica count and cloud bill are things you can move.
• The team is seven people, so you'd have room to set direction on reliability. The stack already exists and works, so you're not starting from a blank page.

What You'll Do

• Own reliability across the production stack: prediction serving, the bidding engine, pacing, delivery. Define the SLOs, build the observability to see them, and shape how we run on-call.
• Bring p99 down on the serving path and run the bidding engine on fewer replicas without giving up latency.
• Work on the data layer under real-time serving. Redis comes first: access patterns, caching strategy, behavior when it degrades.
• Build a way to stress-test the platform end to end in staging so we find our ceiling before production does.
• Handle capacity planning and infrastructure cost as advertiser volume grows.
• Work closely with the data scientists shipping models into this stack, and with Benoît Delsart, who covers the platform and tooling side.
• Run incident response and make sure postmortems turn into changes that ship.
• Collaborate with other engineering teams, notably DSP, since our systems interact heavily with the rest of the stack.

What You Need

• You've owned latency-sensitive services in production, on-call included. You've been paged for something serious and know how that goes.
• Solid Kubernetes and cloud fundamentals. At ease debugging a p99 latency problem
• Observability experience: metrics, tracing, logs, and opinions about what makes an alert worth waking someone for.
• Infrastructure as code, CI/CD. You can ship the fix, not just find it.
• You're comfortable owning systems you didn't design. You'll be on the hook for this stack before you understand every model in it.

Nice to Haves

• High-QPS, low-latency serving experience: ad tech, trading, gaming, streaming or similar.
• Debugging depth on Redis, Kafka or similar.
• Capacity planning and cost work on a large footprint.
• Experience with model rollouts and why they differ from service rollouts.
• Background in programmatic, auctions or CTV.
• Paris-based strongly preferred. We support relocation/visa sponsorship.
• Experience with Go, Rust, Python

What We Offer

For candidates in France, the pay range for this role is €70,000-€110,000.

• Variable pay — Based on real Vibe business goals.
• Hybrid flexibility — We're in the heart of Paris and our team is in 3x a week!
• Health insurance — Full coverage via Alan.
• Meal vouchers — Via Swile.
• Annual offsite — The whole team, once a year, somewhere worth the trip.
• Tech Syncs — Engineering and Product meet in person at least quarterly, worldwide.

Our Interview Process

• Recruiter screen — 30 min. We'll share context; you'll share yours.
• Manager interview — Meet Clément, the person you'd work with directly.
• Technical Debugging and Systems Design Interview — with Performance engineers
• Calibration interview — A senior leader joins to ensure we're holding a consistent, high bar.
• Offer — Fast. We don't let good decisions sit.
• Reference checks — Two calls, handled with discretion.