Nina H.

U.S. job search challenge update: What actually happens in an ML Engineer interview? 👇 After my last post, a lot of people asked two things: 1. “How do I land interviews?” 2. “And what should I expect once I get one?” So today I’m breaking down the most common ML Engineer interview flow I’ve seen across U.S. companies (it may vary by team, but the structure is usually similar): 1️⃣ Recruiter / HR Screen (20–30 min) Goal: confirm you match the role. Keep it simple: your stack, projects, impact, and why this position 2️⃣ Live Coding (45–60 min) Usually LeetCode easy/medium. You’re evaluated not only on the solution. But on how you think out loud, communicate, and test your approach 3️⃣ ML/AI Technical Interview Core ML topics + what’s relevant for the role (e.g., NLP, CV, recommender systems). Expect questions on model selection, evaluation, overfitting, data leakage, metrics, and trade-offs 4️⃣ ML System Design “How would you build this end-to-end?” From problem → data → training → serving → monitoring → iteration. This is where many candidates struggle because it’s not just theory. it’s real-world thinking 🧠 5️⃣ Behavioral Interview The “unusual situations” part 😅 They want to see how you handle conflict, ambiguity, deadlines, mistakes, and teamwork. Fast and clearly 6️⃣ Final Interview (Hiring Manager / Team Lead) Alignment check: what you’ll do day-to-day, team expectations, and sometimes compensation talk • • • If you’re also interviewing in ML/AI (or hiring) - drop a comment: which step is the hardest for you?🔥 #MachineLearning #AI #MLEngineer #InterviewPrep #JobSearch #CareerJourney

19

2 Comments

I built a recommendation system twice. Both times it worked. Both times nobody used it. Second time around, I was demoing it and a user asked: "But why did it recommend this person?" I didn't have an answer. Just metrics. Turns out they didn't want recommendations. They wanted to search and understand why someone was a match. I'd spent months building the wrong thing. Now I try to ask who actually has this problem and what changes if I solve it, before writing any code. Sounds obvious, but it wasn't to me. Wrote about this in my first article in Substack.

51

4 Comments

For our Data Science assignment, Muhammad Aamir and I spent the past week building HSRIS — a Hybrid Semantic Retrieval & Intelligence System — entirely from scratch on ~8,470 real customer support tickets. No sklearn. No HuggingFace. Just base PyTorch, NumPy, and a lot of debugging sessions. ─── What's inside the system ─── ↳ Custom Label & One-Hot Encoders with unseen category handling ↳ Regex tokenizer + top-5,000 vocab + bigram/trigram generator ↳ TF-IDF built manually with smooth IDF — stored as torch.sparse (critical: dense storage crashes Kaggle RAM at 8,470 × 5,000) ↳ GloVe 300d loaded into torch.nn.Embedding ↳ TF-IDF weighted sentence vectors — prevents semantic dilution ↳ Hybrid search: FinalScore = 0.4 × TF-IDF + 0.6 × GloVe ↳ Batch similarity across Dual T4 GPUs via DataParallel ↳ Live Streamlit app with adjustable α slider ↳ Mean Precision@5 = 0.1990 across 200 test queries ───What actually surprised me ─── The math was the easy part. The dataset was full of noise — unresolved template placeholders, log fragments, timestamps — that made even correct retrieval look broken. Cleaning that was the real engineering work. The moment that made it click: querying "money transfer issue" and watching GloVe correctly retrieve "billing problem" tickets that TF-IDF completely missed. Semantics > keywords. ─── Links ─── 🔗 Live App: https://lnkd.in/dRxm3m7m 📓 GitHub: https://lnkd.in/dY34DPfm ✍️ Medium: https://lnkd.in/dS5RrRiZ Thanks to Muhammad Aamir for being a solid partner through every late-night debugging session on this one. #NLP #PyTorch #MachineLearning #GloVe #SemanticSearch #DataScience #Kaggle #DeepLearning #StudentProject #FAST

35

2 Comments

🚀 Excited to share my project: Smart Persian Product Search! Finding the right product on Persian e-commerce platforms is often challenging due to: 🔸 Different naming conventions 🔸 Inconsistent attributes like color, size, and material 🔸 Traditional search engines failing to combine semantics with structured attributes To tackle this, I built a semantic & attribute-based search engine that integrates: ✨ Semantic Search – Sentence Transformers for query understanding ✨ Attribute Matching – gender, size, color, material, subcategory ✨ Keyword Boosting & Scoring – TF-IDF for term relevance ✨ Synonym Mapping – handling Persian variations ⚡ FAISS Indexing – for fast similarity search 🛠 FastAPI REST API – seamless integration with e-commerce platforms This system helps Persian online shops deliver faster, more accurate product discovery, even with incomplete or ambiguous queries. 🔗 Check out the full project on my GitHub

27

After over a decade in ML hiring, I’ve seen a pattern: Many interview loops reward confidence over competence. When hiring outcomes feel inconsistent, the root cause is often weak signal design. Three questions usually expose it: 1) Can every interviewer state what they measured, in the same terms? If not, you’re likely collecting opinions rather than signals. 2) Does each round reduce uncertainty in a distinct way? If two rounds test the same thing then fatigue builds, not breadth of signal. 3) Can you explain a “no” without relying on feelings? If feedback is gut-based then the signal wasn’t clear. I’ve started posting again to share my thoughts and insights for building reliable hiring signals in ML (Applied Science, MLE, and Enablement) roles, and to examine where AI tools help, and where they don’t. I'm curious as to what’s the hardest thing for your teams to assess in ML interviews right now?

17

A few years ago, I was applying to roles almost every day. Tailoring resumes, grinding through mock interviews, building projects, refreshing Gmail hoping something changed. Doing everything you’re supposed to do. And still, silence. Or rejections with no real explanation. What made it harder was being an international student. Every application felt heavier than just a job application. There were visa timelines, loan pressure, and a quiet voice asking: what happens if this doesn’t work out in time? For a while, I genuinely thought I just wasn’t good enough. It took time to realize that wasn’t it. The market was brutal for almost everyone post-COVID. A lot of the people I knew going through the same thing are doing really well now, which tells you something. But what actually changed things wasn’t working harder. It was being more intentional. About how I was showing up, who was actually seeing my work, and whether the effort was going in the right direction. I’m still learning, and I’m not fully where I want to be. But those early years shaped how I think about careers now. Over the next few posts I want to share some of what I picked up, not as someone who has it all figured out, but as someone a few steps further down the same road. If you’re early in your career or going through a tough search right now, this one’s for you🙌🏻 #JobSearch #InternationalStudents #TechCareers #GenerativeAI #CareerGrowth

47

2 Comments

Last year, a hiring manager told me something that completely changed how I talk about my ML projects: “I don’t care how smart your model is if my team’s week still feels the same.” At first, it stung. But the more projects I shipped, the more it made sense. Most of us present our work like this: • “Built churn model with 0.9 AUC” • “Improved accuracy by 5%” • “Implemented recommendation system with XYZ algorithm” Meanwhile, the people who hire us are wondering: • “Will this person reduce the stress in my team?” • “Will my PMs and business stakeholders finally get clearer decisions?” • “Will we trust our dashboards more next quarter than we do today?” On one project, the metric wasn’t AUC at all. It was: “Can the team stop dreading Monday reports?” What made the difference wasn’t a new architecture. It was sitting with the team, understanding where their week actually hurt, and then using Python + SQL + some basic ML to give them back a few hours and a lot of confidence. That’s what they remembered in my performance review—not the algorithm name. If you’re a recruiter, hiring manager, or business lead working with data teams: What’s one question you ask to figure out whether someone can change how your week feels, not just your metrics? #datascience #machinelearning #analytics #hiring #recruiting #businessimpact

10