Priya Mehta

Layoffs = Growth? Bumble just proved it. 🚨 Bumble’s stock soared 26% overnight — not because of a new product, not because of a market expansion… …but because they laid off 30% of their workforce. That’s ~240 jobs cut to save $40 million/year. Harsh? Yes. Strategic? The market sure thinks so. 📉 Bumble’s valuation had plummeted from $7.7B in 2021 to just $538M. Now, with founder Whitney Wolfe Herd back in the CEO seat, they’re rewriting the playbook: → Slim down. → Reinvest in product and tech. → Focus on execution, not just expansion. It’s a cold reminder of where public markets place their bets: efficiency and clarity of vision > bloated org charts. 💬 This isn’t just a Bumble story. It’s a broader signal across tech: In 2025, lean may be the new scale.

72

19 Comments

Right now, it’s not just H1B holders under pressure, the ones really feeling the burn are 𝘀𝘁𝘂𝗱𝗲𝗻𝘁𝘀, 𝗳𝗿𝗲𝘀𝗵 𝗴𝗿𝗮𝗱𝘂𝗮𝘁𝗲𝘀 𝗮𝗻𝗱 𝗮𝗹𝗹 𝘁𝗵𝗲𝘀𝗲 𝗨𝗦 𝗳𝗼𝗹𝗸𝘀 𝗮𝗿𝗲 𝗰𝗼𝗺𝗶𝗻𝗴 𝗯𝗮𝗰𝗸 𝘁𝗼 𝗲𝗮𝘁 𝗜𝗻𝗱𝗶𝗮𝗻 𝗜𝗧 𝗷𝗼𝗯𝘀. So, the bigger hit is actually on:  • Students studying here in US with crazy loan amounts on their head.  • Folks who graduated but still don’t have a job. Many students are already stressed about getting a job, and now this adds on. Dealing with rent, family pressure, loan interest, and living expenses all at once, so just imagine. And don’t think this stops in the US, it’s gonna bounce back to India too. The first impact? 𝗜𝗻𝗱𝗶𝗮𝗻 𝗜𝗧. Bcz the people heading back from the US will be looking at those jobs, and let’s be real, they already have an edge. They’ve worked with 𝗠𝗔𝗔𝗡𝗚/𝗙𝗔𝗔𝗡𝗚, on 𝗿𝗲𝗮𝗹 𝗽𝗿𝗼𝗱𝘂𝗰𝘁-𝗯𝗮𝘀𝗲𝗱 𝗽𝗿𝗼𝗷𝗲𝗰𝘁𝘀, working on 𝗹𝗮𝘁𝗲𝘀𝘁 𝘁𝗲𝗰𝗵 𝗹𝗶𝗸𝗲 𝗔𝗜/𝗠𝗟. Sure, India has a lot of potential and a big gap to fill in tech. But the real question is, 𝗮𝗿𝗲 𝘄𝗲 𝗿𝗲𝗮𝗱𝘆 𝘁𝗼 𝘁𝗮𝗸𝗲 𝗶𝗻 𝘁𝗵𝗶𝘀 𝘄𝗮𝘃𝗲 𝗼𝗳 𝗿𝗲𝘁𝘂𝗿𝗻𝗲𝗲𝘀? Truth is, 𝘁𝗵𝗲𝘆’𝗹𝗹 𝗲𝗮𝘀𝗶𝗹𝘆 𝗼𝘂𝘁𝗰𝗼𝗺𝗽𝗲𝘁𝗲 𝗮 𝗹𝗼𝘁 𝗼𝗳 𝘁𝗵𝗲 𝗲𝘅𝗶𝘀𝘁𝗶𝗻𝗴 𝘁𝗮𝗹𝗲𝗻𝘁 𝗶𝗻 𝗜𝗻𝗱𝗶𝗮. Why? Bcz most niche technologies emerge first in the US and only later get outsourced to Indian firms like Infosys or TCS. 𝗦𝗼𝗺𝗲 𝗵𝗮𝗿𝘀𝗵 𝗿𝗲𝗮𝗹𝗶𝘁𝗶𝗲𝘀:  • 𝗗𝗼𝗻’𝘁 𝗿𝘂𝘀𝗵 𝘁𝗼 𝘁𝗵𝗲 𝗨𝗦 𝗿𝗶𝗴𝗵𝘁 𝗻𝗼𝘄. There are already too many students stuck in the job hunt, some graduated, some still studying, and many more on the way.  • The numbers are insane, 𝗼𝘃𝗲𝗿 𝟭.𝟭𝟮 𝗺𝗶𝗹𝗹𝗶𝗼𝗻 international students in the US (2023-24, record high), 𝟯𝟯𝟭𝗸+ from India—𝟮𝟯% jump in a year. Another report puts Indians at 𝟰𝟮𝟬𝗸 (𝟮𝟳% of all students). Massive pressure on the market.  • A big reason? 𝗨𝗻𝗶𝘃𝗲𝗿𝘀𝗶𝘁𝗶𝗲𝘀 𝘄𝗮𝗶𝘃𝗲𝗱 𝗚𝗥𝗘 𝗮𝗻𝗱 𝗧𝗢𝗘𝗙𝗟/𝗜𝗘𝗟𝗧𝗦 during covid, which made it way easier for people to get in US, even if they weren’t really ready for the competitive job market here.  • On top of that, 𝘀𝗶𝗻𝗰𝗲 𝗖𝗵𝗮𝘁𝗚𝗣𝗧 𝗯𝗹𝗲𝘄 𝘂𝗽 𝗶𝗻 𝟮𝟬𝟮𝟯, 𝗰𝗼𝗺𝗽𝗮𝗻𝗶𝗲𝘀 𝗲𝘃𝗲𝗿𝘆𝘄𝗵𝗲𝗿𝗲 𝗰𝗵𝗮𝗻𝗴𝗲𝗱 𝗰𝗼𝘂𝗿𝘀𝗲 𝗮𝗻𝗱 𝘄𝗲𝗻𝘁 𝗮𝗹𝗹-𝗶𝗻 𝗼𝗻 𝗔𝗜. Now tell me, 𝗵𝗼𝘄 𝗺𝗮𝗻𝘆 𝘂𝗻𝗶𝘃𝗲𝗿𝘀𝗶𝘁𝗶𝗲𝘀 𝗶𝗻 𝗜𝗻𝗱𝗶𝗮 𝗮𝗿𝗲 𝘀𝗲𝗿𝗶𝗼𝘂𝘀𝗹𝘆 𝗯𝘂𝗶𝗹𝗱𝗶𝗻𝗴 𝗔𝗜/𝗠𝗟-𝘀𝗸𝗶𝗹𝗹𝗲𝗱 𝗴𝗿𝗮𝗱𝘀? And even if they are, 𝗵𝗼𝘄 𝗺𝗮𝗻𝘆 𝗰𝗼𝗺𝗽𝗮𝗻𝗶𝗲𝘀 𝗮𝗿𝗲 𝗮𝗰𝘁𝘂𝗮𝗹𝗹𝘆 𝘄𝗼𝗿𝗸𝗶𝗻𝗴 𝗼𝗻 𝗻𝗶𝗰𝗵𝗲 𝗔𝗜 𝗽𝗿𝗼𝗷𝗲𝗰𝘁𝘀 𝗶𝗻 𝗵𝗲𝗮𝗹𝘁𝗵𝗰𝗮𝗿𝗲, 𝗳𝗶𝗻𝗮𝗻𝗰𝗲, 𝗼𝗿 𝗼𝘁𝗵𝗲𝗿 𝘀𝗲𝗰𝘁𝗼𝗿𝘀 on their own? Not many. So yeah, 𝘁𝗵𝗶𝘀 𝘄𝗮𝘃𝗲 𝗶𝘀 𝗴𝗼𝗶𝗻𝗴 𝘁𝗼 𝗵𝗶𝘁 𝗜𝗻𝗱𝗶𝗮𝗻 𝗜𝗧 𝘁𝗵𝗲 𝗵𝗮𝗿𝗱𝗲𝘀𝘁. And if anyone in India is feeling happy about US layoffs or immigration issues, 𝗷𝘂𝘀𝘁 𝗯𝗲 𝗰𝗮𝗿𝗲𝗳𝘂𝗹, 𝗯𝗰𝘇 𝗶𝗳 𝘁𝗵𝗲𝘀𝗲 𝗨𝗦 𝗳𝗼𝗹𝗸𝘀 𝗮𝗿𝗲 𝗰𝗼𝗺𝗶𝗻𝗴 𝗯𝗮𝗰𝗸... 𝗮𝗻𝗱 𝘁𝗵𝗲𝘆’𝗿𝗲 𝗰𝗼𝗺𝗶𝗻𝗴 𝗳𝗼𝗿 𝘆𝗼𝘂𝗿 𝗷𝗼𝗯𝘀.

111

18 Comments

Building an LLM - concepts and interview question series Topic - Self-Attention If you have seen my profile header you know the Vaswani et al paper is one of my favorites but often in interviews when asked to give an intuition about "How transformers work?" - the candidates jump straight to the equation (formula 1 in the pic). Attention is a relevance filter. Imagine every token in a sentence is looking at every other token and asking " Hey, how much should i care about you?" When the model processes a word say "bank", it looks around neighbors like "river" or "money" to decide which definition holds true. This looking around is calculated via Attention scores. To compute this mechanically, we use three vectors - Query (Q): What the current token is looking for (e.g. "I am a verb, looking for my subject"). Key (K): What every other token identifies as (e.g. "I am a noun/subject"). Value (V): The actual content/meaning of the token. The "Attention" is just the match between the Query and the Key. If they match well, the model pays attention and absorbs the Value. This would also explain why Large Language Models are so good at long-range context and interpretation compared to previous architectures. 2 very good resources I recommend for understanding this better 1) Building LLMs from Scratch by Sebastian Raschka, PhD - Chapter 3 Coding Attention Mechanism. He also has a Youtube series on this but the book chapter is much more detailed. 2) Youtube tutorial by 3Blue1Brown - The animations are the best that I have seen in explaining attention. Follow up Question - Why is the square root of dk required in the Attention calculation? Answer in comments. Happy Learning #LLMs #GenAI #DataScience #MachineLearning #interview

159

7 Comments

Tiger Analytics is known for asking some solid, real-world Data Engineering questions during interviews. If you're preparing, here’s a quick snapshot of what they typically expect 👇 💰 Role-wise Compensation (Approx.) Data Engineer: ₹15–20 LPA Senior Consultant – Data Engineering: ₹25–30 LPA 📌 Commonly asked technical questions: 1️⃣ How would you build a highly scalable ingestion pipeline for hourly multi-TB parquet files landing in ADLS using ADF + Databricks? Consider partitioning, pruning, and metadata performance tuning. 2️⃣ Explain the internals of OPTIMIZE and ZORDER in Delta Lake. Also discuss the scenarios where running them may not be beneficial. 3️⃣ Write a SQL query to find customers who purchased in 3 or more back-to-back months. 4️⃣ What’s the performance difference between using Delta Lake CDF and MERGE? When is one a better choice over the other? 5️⃣ How does Spark’s Tungsten engine speed up execution? Talk about off-heap memory and whole-stage code generation. 6️⃣ A Databricks job fails intermittently due to data skew. How do you diagnose it and what steps would you take to mitigate it? 7️⃣ What are the best practices for building resilient ADF pipelines? Think retries, monitoring, alerting, and custom logging. 8️⃣ How would you design a watermark strategy in Spark Structured Streaming when events arrive late or out of order? 9️⃣ Delta Live Tables vs traditional Spark notebooks: When should each be used for production-grade ETL? 🔟 Explain SCD Types 1, 2, and 3. How would you efficiently implement Type 2 using Window functions + MERGE? 1️⃣1️⃣ How do you build enterprise data lineage with Purview integrated across ADF and Databricks? 1️⃣2️⃣ Compare Databricks SQL Warehouse with Azure Synapse Dedicated SQL Pool for analytics workloads. 1️⃣3️⃣ Describe a real scenario where you applied data quality checks (schema, nulls, referential integrity) and how you set up automated alerts. 1️⃣4️⃣ Share an experience where you reduced Databricks costs using autoscaling, spot nodes, or better query optimization. 1️⃣5️⃣ Design a near real-time Kafka → Spark Streaming → Delta pipeline. How do you guarantee exactly-once processing and handle replay logic? 👉 If you found this helpful, hit LIKE 👍 and RESHARE 🔁 so more people can benefit from it! 📍 𝗙𝗼𝗹𝗹𝗼𝘄 𝗺𝗲 𝗵𝗲𝗿𝗲 𝗳𝗼𝗿 𝗱𝗮𝗶𝗹𝘆 𝗦𝗤𝗟, 𝗗𝗮𝘁𝗮 𝗘𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝗶𝗻𝗴, 𝗮𝗻𝗱 𝗝𝗼𝗯 𝘂𝗽𝗱𝗮𝘁𝗲𝘀: 👉 https://lnkd.in/dhYpfrS7 𝗜 𝗵𝗮𝘃𝗲 𝗽𝗿𝗲𝗽𝗮𝗿𝗲𝗱 𝗖𝗼𝗺𝗽𝗹𝗲𝘁𝗲 𝗜𝗻𝘁𝗲𝗿𝘃𝗶𝗲𝘄 𝗣𝗿𝗲𝗽𝗮𝗿𝗮𝘁𝗶𝗼𝗻 𝗚𝘂𝗶𝗱𝗲 𝗳𝗼𝗿 𝗗𝗮𝘁𝗮 𝗘𝗻𝗴𝗶𝗻𝗻𝗲𝗿𝘀/𝗗𝗮𝘁𝗮 𝗔𝗻𝗮𝗹𝘆𝘀𝘁𝘀. 👉𝗚𝗿𝗮𝗯 𝗶𝘁 𝗵𝗲𝗿𝗲(𝟭𝟬𝟬% 𝗥𝗲𝗳𝘂𝗻𝗱): https://lnkd.in/giHUMNx9

110

4 Comments

8 years ago, I was writing Hive queries and tuning Spark jobs on Hadoop clusters. Today, I'm building Multi-Agent RAG systems that help wealth management advisors make compliance-grounded decisions in real time. The jump from Data Engineer → Data Scientist → AI/ML Engineer wasn't a straight line. It was a series of bets on what mattered next. Here's what I've learned building production AI across Healthcare, Financial Services, and Insurance: 🔹 Predictive ML alone isn't enough anymore. The real unlock is when your churn model feeds directly into an LLM prompt — combining probability scores with document-backed reasoning. That's where explainability actually lives. 🔹 LangGraph > LangChain for anything compliance-sensitive. Linear chains break under audit requirements. Explicit agent nodes with checkpointing and retry logic don't. 🔹 RAG is only as good as your retrieval strategy. FAISS in dev, Pinecone in prod, and obsessing over top-K thresholds is what separates a demo from a deployment. 🔹 MLOps isn't optional. Drift detection, PSI scoring, audit logging, RBAC — these aren't enterprise checkbox items. They're what keeps your model trusted 6 months after launch. The most underrated skill in AI engineering? Knowing when NOT to use an LLM. Building hybrid systems — where deterministic ML handles the signal and LLMs handle the reasoning — is where the real value gets created. Curious what patterns others are seeing in production GenAI systems. What's the hardest thing to get right at scale? #GenerativeAI #MachineLearning #RAG #LLMs #MLOps #AIEngineering #LangChain #FinancialAI #HealthcareAI

6

The majority of Data Science candidates I interview have one project in common: Clustering. Be it, Customer segmentation. User behavior grouping. Product usage patterns. Yet most candidates struggle the moment the questions go beyond “I used K-Means.” Here are some conceptual clustering questions that can genuinely help you in your next Data Science interview — with explanations. 1. Why does K-Means fail even when you choose the correct number of clusters (K)? K-Means assumes: Clusters are spherical All clusters have similar variance The mean is a good representation of the cluster center In real-world data, clusters are often: Elongated Unevenly dense Influenced by outliers When these assumptions break, K-Means produces misleading clusters — even if K is technically correct. Interview insight: Mention assumptions first. That shows maturity. 2. Why is feature scaling critical before applying K-Means? K-Means relies on Euclidean distance. If features are on different scales: Revenue dominates Age Transaction count dominates tenure The algorithm clusters based on scale, not meaning. Fix: Standardization or normalization Or switch to algorithms less sensitive to distance assumptions Key takeaway: Distance-based models amplify bad preprocessing. 3. How is K-Modes fundamentally different from K-Means? Most people say: “K-Modes is for categorical data.” That’s incomplete. Real difference: K-Means minimizes squared distance from the mean K-Modes minimizes category mismatches from the mode K-Modes uses: Frequency-based calculation Category agreement instead of geometry Why it matters: Means don’t exist for categorical variables. Modes do. 4. What problems arise when using K-Modes with high-cardinality features? High-cardinality categories (like user IDs or zip codes) can: Introduce noise Reduce interpretability Common fixes: Group rare categories Drop non-informative features Use feature weighting or K-Prototypes for mixed data Interview signal: Talk about feature relevance, not just algorithms. 5. Why does Euclidean distance perform poorly for time-series clustering? Two time-series can: Have identical patterns Be shifted in time Euclidean distance treats them as very different. Example: Seasonal sales curves shifted by one month. Solution: Dynamic Time Warping (DTW) Shape-based distance measures 6. How do you decide the number of clusters when the elbow method is unclear? This happens more often than not. Use: Silhouette score Stability across multiple runs Business interpretability Reality check: K is rarely “found.” It is chosen with context. 7. How do you validate clustering results without labels? Since clustering is unsupervised: Evaluate separation and compactness Check consistency across samples Analyze feature distributions per cluster Validate usefulness with stakeholders If clusters don’t change decisions, they’re just visuals. Follow for more interesting information in the data world. #DataScience #Interview

109

9 Comments