Samsung · Primly Community

Samsung machine learning engineer interview: what the loop covers and where I got tripped up

ml_mike · 5 replies

Did the Samsung MLE loop in Q1 2026, targeting a role on their on-device AI team (think Bixby, NPU optimizations, that sphere). Posting because most Samsung ML interview content you find is either for research scientist roles or totally generic.

Background: 8 years, mostly NLP and recsys, now leaning into inference optimization.

The process. Recruiter screen, technical phone screen, then a 5-round virtual onsite (they spread it across two half-days). One round was a take-home, but they gave the option to skip it if you did well on the phone screen. I skipped it.

Technical phone screen. 45 minutes. Part 1: ML fundamentals. Explain the bias-variance tradeoff, how gradient descent finds a minimum, when would you use L1 vs L2 regularization and why. Part 2: a short Python coding exercise, nothing algorithmic, more like 'given this NumPy array, compute X.' They're checking if you can code, not if you've memorized LeetCode.

Onsite round 1 (ML system design). Design a recommendation system for app discovery on a Galaxy device. I walked through the full stack: data collection, feature engineering, candidate generation (two-tower model), ranking model, serving infrastructure, A/B testing loop. They pushed hard on latency constraints because on-device matters to them. Talk about model quantization, INT8 inference, and TensorRT or equivalent if on-device is in the JD.

Onsite round 2 (deep technical). This is where I stumbled. They wanted to talk through backprop by hand and then pivot into transformer architecture details. Self-attention complexity (O(n^2) in sequence length), how you'd speed that up (sparse attention, linear attention), and why KV cache matters for inference. I knew this but got flustered on the notation. Go in with this fresh.

Onsite round 3 (applied ML coding). Given a dataset, write a simple training loop in PyTorch. Not complex, but you need to know the basics cold: DataLoader, optimizer, loss.backward(), scheduler. They're not testing whether you can build GPT, they're testing if you can write real code.

Rounds 4 and 5: Behavioral and a "vision" round with the director. The vision round was actually interesting. They asked what I think the ML landscape looks like in 3 years, specifically for edge and on-device inference. Have a real answer.

Offer for senior MLE was around $165k base + bonus. Equity in Samsung stock, not options. Total comp was workable but not hyperscaler territory. The work on the NPU side is genuinely interesting if that's your thing.

5 replies

ux_uma

The backprop by hand question is something I always dread. How detailed did they get? Like do they want you to actually derive the chain rule through an attention layer or just explain conceptually?

ml_mike

Somewhere in between. They had me write out the gradient flow through a simple 2-layer network by hand (not attention specifically), then asked conceptually how you'd extend it. You don't need to derive multi-head attention gradients symbolically but you do need to understand where the computation graph gets complicated.

sec_sasha

The on-device latency angle is so specific to Samsung and it's exactly what people don't prep for. If you're interviewing there for anything AI-adjacent, understanding the NPU pipeline and why INT8 quantization matters is table stakes.

returner_ren

How receptive were they to candidates coming from more traditional ML backgrounds (not specifically on-device)? I have recsys experience but nothing embedded.

ml_mike

My recruiter was clear that they have teams in both areas. The on-device team is niche but they also hire for cloud-side AI that powers SmartThings, Samsung Health, etc. Ask in the recruiter screen which org the role sits in. That changes the prep completely.