My Experience Interviewing with Huawei Vancouver for an ML Research Role

What Was Promised: The ML Research Dream Job

Let’s be clear — the job ad was good. Like, really good. “Senior ML Research Scientist,” Vancouver-based, focus on multimodal understanding, natural language generation, and federated learning applications in telecom. Salary: $130K–$160K CAD. Remote options. PhD required or equivalent experience. Bonus points for NeurIPS or ICML publications.

I’ve been in tech for eight years. I know how to read between the lines. This wasn’t a “ML engineer to maintain recommendation models” role. This was pitched as genuine research — the kind where you file patents, co-author papers, and attend conferences. The team was supposedly a bridge between Huawei’s Beijing AI lab and North American telecom partners. Sounded like a chance to work on real problems with real impact.

And yeah, I was intrigued. I’ve got a background in NLP and spent two years at a fintech startup building transformer-based fraud detection. Not Google Brain, but solid. My plant factory in Icheon? I’ve been experimenting with vision models to detect lettuce stress from RGB images. Small scale, but proof I still code.

The Job Posting That Lured Me In

The posting was on LinkedIn and Huawei’s career site. Clean layout. Listed skills: PyTorch, distributed training, federated learning, experience with large-scale datasets. Bonus: knowledge of 5G network constraints (interesting twist). The “Responsibilities” section mentioned “publishing findings in top-tier conferences” and “collaborating with international teams.” That’s the stuff research folks live for.

What stood out? No mention of SDLC, Agile, or CI/CD. No “own the full stack.” Just pure research language. I applied the same day.

Why Huawei Vancouver Seemed Like a Legit Research Hub

Look — Huawei has a complicated reputation in the US. Security concerns, export bans, the Meng Wanzhou thing. But their research output? Undeniable. They file thousands of patents a year. Their 2023 AI paper count was top 5 globally. And Vancouver has become a quiet tech hub for Chinese firms setting up R&D outposts. Baidu, Tencent, Huawei — all have offices there.

I reached out to a former grad school buddy who worked at Huawei Beijing for six months. He said the research side was real, but the pressure to deliver for product teams was intense. “They publish, but it’s often incremental,” he said. Still, I figured Vancouver might be different — more autonomy, less product pressure.

What Actually Happened: The Interview Reality Check

First contact came three weeks after applying. Recruiter was friendly, confirmed the role, sent a calendar invite. Two 1-hour technical rounds, one 45-minute HR chat, then a final panel. Standard.

Then came the prep docs. That’s when I got my first clue.

Round 1: Coding Grindhouse

The first interview was with a senior software engineer. Title: “ML Research Scientist.” But the focus? LeetCode medium-to-hard. We did three problems in 60 minutes:

• Reverse a linked list (iterative and recursive)
• Find the longest palindromic substring
• Implement a thread-safe LRU cache

Nothing was ML-related. Not even a softmax implementation. No discussion of loss functions, embeddings, or model tradeoffs. It felt like I was interviewing for a backend role at Amazon.

I passed — barely. I hadn’t touched LRU caches since my 2016 Google interview prep. But I coded it. Barely.

Round 2: System Design That Felt Like a Backend Trap

Second round: system design. “Design a distributed key-value store.” Classic. Consistency models, sharding, replication, failover. I’ve done this before. But again — zero connection to machine learning.

The interviewer kept pushing on CAP theorem tradeoffs and Paxos vs. Raft. Cool topics, but not what I’d expect for a research role. When I asked, “How does this tie into the team’s work on federated learning?” he said, “Oh, that’s for the next round.”

And yeah, I nailed the design. But I left wondering: is this what they really care about?

Round 3: Research Discussion? Not Really

The final round was supposed to be with the research lead. Instead, it was two engineers and a manager. They asked about my PhD work — for five minutes. Then pivoted to: “Walk us through how you’d deploy a model on an edge device with 200MB RAM.”

I gave a solid answer — quantization, pruning, ONNX runtime, maybe TinyML. But then they asked: “What if the device loses power mid-inference? How do you checkpoint?” That’s an embedded systems question, not a research one.

They never asked about my papers. Never discussed open problems in multimodal learning. No interest in my plant factory vision model. Nothing.

And the kicker? At the end, the manager said, “We’re looking for someone who can go from research to production quickly.” Translation: we want a full-stack ML engineer, not a researcher.

The Mismatch: Research Role vs. SWE Evaluation

Here’s the thing: if you’re hiring for a research scientist, evaluate research. Not LeetCode speed.

I’ve sat on hiring committees. I know how this goes. Research roles should test:

• Depth of understanding in core ML concepts
• Ability to critique and extend existing work
• Experimental design and evaluation rigor
• Communication of complex ideas
• Code quality in research contexts (notebooks, training scripts, etc.)

Not whether you can crank out a red-black tree from memory.

When PhDs Are Judged Like New Grads

I’ve been on both sides. When I was hiring at my last startup, we had two tracks: research and engineering. For research candidates, we gave a paper to read and asked them to critique it. Then a 30-minute discussion. For engineers? Yeah, we did coding challenges.

Blending the two is lazy. It’s also wasteful. You end up filtering out people who’ve spent years thinking deeply about models, just because they forgot how to balance a binary search tree.

And let’s be real — most senior researchers haven’t touched Dijkstra’s algorithm in a decade. That’s fine. They’re building systems, not taking algorithms final exams.

Why This Kind of Interview Fails ML Candidates

The problem isn’t just unfairness. It’s misalignment.

If you test ML researchers on SWE skills, you’ll hire SWEs who dabble in ML. That’s not the same.

I’ve seen this before. A team claims they’re doing “AI research,” but really they’re just fine-tuning Hugging Face models and slapping a Docker container on it. No novel architecture work. No ablation studies. No real research.

And yeah, maybe that’s what Huawei Vancouver is doing. Maybe the “research” label is just to attract PhDs who’ll work for less than FAANG.

Is This How China-Based Tech Firms Operate Abroad?

I don’t know. But I’ve heard patterns. Baidu’s Silicon Valley lab shut down in 2019. Tencent’s Seattle office downsized. Huawei’s Waterloo AI lab was reportedly restructured after 2020 sanctions.

These outposts often start with big promises. Then reality hits: parent company demands product alignment, budgets tighten, research gets deprioritized.

Maybe the Vancouver team used to do real research. Now? Feels like they’re just supporting product teams with model deployment.

Alternatives That Actually Value Research

Don’t get me wrong — there are still places doing real ML research.

They’re just not always the obvious ones.

Companies That Still Do Real ML Research

👉 Best: michigan-farm-town-voted-down-plans_02121794236.html” class=”auto-internal-link”>Microsoft Research (Redmond, Montreal, New England). Still publishes at top venues. Still hires PhDs to work on open problems. Yes, some teams are product-aligned. But MSR has autonomy.

Google Brain and DeepMind? Still solid, but harder to get into. And more politicized. Meta AI (FAIR)? Hit or miss. Some teams do groundbreaking work. Others are just tuning Llama for ads.

Amazon? Almost entirely product-driven. Their research is often just a thin wrapper over existing work.

Smaller Labs With Academic Freedom

Here’s where it gets interesting.

Elemental AI (acquired by ServiceNow) used to be great. Now? Gone.

But places like Mila in Montreal, Vector Institute in Toronto, and Amii in Edmonton are still doing strong work. They partner with industry but maintain independence.

And they actually interview researchers like researchers. Paper discussions. Project walkthroughs. No LeetCode.

Hybrid Roles That Balance Code and Papers

If you want to publish and ship code, look at:

• NVIDIA Research — strong in computer vision and systems
• Intel Labs — underrated, still does hardware-ML co-design
• Allen Institute for AI (AI2) — fully research-focused, based in Seattle

These places understand that research isn’t just another sprint task.

Red Flags to Watch For in ML Job Postings

After my experience interviewing with Huawei Vancouver for an ML research role: strong mismatch between how it was pitched and how it was evaluated [D], here’s what I’ve learned to spot:

Vague Research Scope, Heavy Coding Tests

If the job ad says “publish papers” but the interview process is all LeetCode, run.

Ask: “Will I be expected to write production-grade code daily?” If yes, and you want research, it’s a mismatch.

No Mention of Publications or Conferences

Real research teams list recent papers. They mention NeurIPS, ICML, or ACL. They’ll say “we encourage conference submissions.” If that’s missing, it’s a red flag.

Interview Panels Without Research Leads

If all your interviewers are software engineers, not researchers, you’re not joining a research team.

Ask to speak with the research lead. If they refuse, that’s your answer.

Frequently Asked Questions

Was the role actually a research position?

No, not in practice. While the job posting claimed it was for ML research, the interview heavily focused on software engineering skills like system design and LeetCode-style coding. There was minimal discussion of research, papers, or innovation. It felt more like a full-stack ML engineer role disguised as research.

Did you get feedback after the interview?

Yes — generic email saying “we’ve decided to move forward with other candidates.” No specific feedback, which is common. When I followed up, the recruiter said the team was looking for “stronger coding fundamentals,” confirming the SWE bias.

Is Huawei Vancouver still hiring for ML roles?

As of June 2024, yes. But based on Glassdoor reviews and my experience, the process remains coding-heavy. Several recent applicants report similar mismatches — research titles, SWE interviews.

Should I avoid applying to Huawei altogether?

Not necessarily. If you want a strong software engineering role with some ML exposure, it could be a fit. But if you’re a researcher looking to publish and explore new ideas, look elsewhere. The culture favors implementation over innovation.

What could Huawei have done better?

Rebrand the role honestly. Call it “ML Engineer” if that’s what they want. Or, if they truly want researchers, evaluate research: ask candidates to discuss papers, design experiments, or extend existing models. Stop using SWE interviews as a proxy for research ability.