Sakshi Sinha

INTERVIEW NO. 3 Organization: Kaiser Permanente Role: Data Scientist - IV Round 1 - The interviewer asked me in detail about my experience in data science and machine learning. The interview lasted for around 45 mins and most of it was focused discussing my previous projects in AI/ML. In the end, he asked to write 1 Python program in which he gave me a list which had multiple duplicate values and asked me to find the occurence of each number along with its index. Round 2 - The interviewer went through my resume and started asking questions based on my profile. He tested my coding skills in SQL and Machine learning by giving me hypothetical scenarios. He asked me to write the code for an end to end ML pipeline for predicting whether a patient is a cancer patient or not. In technical interviews, you need to think and come up with a reasonable solution to the problem that is assigned in limited time. Make sure to work on your critical thinking and problem solving skills. 💡 Interview Tip: Avoid using percentage-based metrics in your resume (e.g., “Saved X% of time through ABC method”) unless you can confidently justify those numbers during the interview. Thanks for reading! Follow #rajreacts #interviewprep #aiml #interviews #datascience #machinelearning

31

1 Comment

The A-Z of Data Engineering I wish someone had told me when I just started my job with Amazon. I love data, and these are 26 observations that have immensely helped in my last 7+ years of career. Some of these even nudged me toward my promotion as a Sr. Data Engineer. If you’re also a fellow engineer/analyst/scientist, drop some nuggets below! ❯ A — Algorithms Start with the simplest algorithm that works. Don’t jump to transformers if logistic regression gets 90% right. ❯ B — Bias Check for class imbalance, over-represented categories, and your own assumptions in feature selection. ❯ C — Cross Validation Use stratified folds for classification. For time series, use walk-forward validation. Don’t just split randomly. ❯ D — Data Cleaning Standardize column names, handle nulls before modeling, and always save pre-clean snapshots for rollback. ❯ E — EDA (Exploratory Data Analysis) Automate what you can (like pandas_profiling) but don’t skip looking at value distributions and outliers manually. ❯ F — Feature Engineering One-hot, label encode, extract date parts, groupby aggregations, they’ll often outperform fancy deep learning. ❯ G — Gradient Descent Understand how learning rate affects convergence. Use learning rate scheduling if your loss plateaus. ❯ H — Hyperparameter Tuning Use RandomizedSearchCV or Optuna. Avoid grid search for large models — it’s wasteful. ❯ I — Imputation Mean/median for numeric, mode for categorical — unless missingness itself is informative, then create an indicator. ❯ J — Jupyter Notebooks Keep them clean. One notebook = one purpose. Push production code into scripts, not cells. ❯ K — K-Means Use it for clustering only when features are normalized and Euclidean distance makes sense. ❯ L — Logging Use MLflow or Weights & Biases to track experiments. Don’t rely on your memory or Excel. ❯ M — Model Versioning Save model files with timestamp + metric. Use DVC or at least folders with good naming conventions. ❯ N — NumPy & Pandas Don’t write for-loops when .apply() or vectorization can do it faster and cleaner. ❯ O — Overfitting Use regularization (L1/L2), cross-validation, and fewer features. Watch for a large gap between train/test. ❯ P — Pipelines Use sklearn.pipeline or make_pipeline to cleanly stack preprocessing and models. Makes deployment easier. ❯ Q — Quantiles Use for outlier detection, feature binning, and robust normalization. Way better than just using means. ❯ R — Random Forest Great baseline model. Use class_weight='balanced' for imbalanced data. Tune max_depth and min_samples_split. ❯ S — SQL Most of your data lives in a database. Learn to write efficient joins, window functions, and subqueries. ❯ T — Time Series Avoid data leakage — don’t mix future values into past. Use Prophet or XGBoost with lag features for quick wins. ❯ U — Unbalanced Data Use SMOTE, class weights, or downsampling. But always compare with the real-world cost of false positives. Continued in comments ↓

205

18 Comments

💡 Real Healthcare Databricks Interview Questions & My Answers (8 Yrs Experience) Sharing 5 real-world scenario-based questions I faced as a Senior Data Engineer in Healthcare 👇 These reflect real production challenges from handling PHI to building scalable Delta pipelines. 1️⃣ How do you ensure HIPAA compliance while processing data in Databricks? 🩺 Answer: “In healthcare projects, I always enforce PHI protection using column-level encryption and access control via Unity Catalog. All PII fields are tokenized before landing in Delta tables, and audit logs are maintained using Databricks audit logs + Azure Monitor. We also integrate with Azure Key Vault for credential rotation and secure secrets management.” 2️⃣ How do you optimize Delta tables storing EHR/Claims data with billions of rows? 📊 Answer: “I apply ZORDER on high-cardinality columns like patient_id or claim_id for faster lookups. Partitioning is done by encounter_date or claim_year to improve pruning. I’ve also used OPTIMIZE with VACPUM scheduling and Delta caching for repeated analytical workloads — achieving a 45% boost in query performance.” 3️⃣ How do you handle incremental updates from HL7/FHIR streaming feeds? ⚙️ Answer: “I built a streaming ingestion using Auto Loader + Structured Streaming. Each micro-batch merges data into the Delta table using MERGE INTO with CDC logic. Watermarking ensures late-arriving updates are captured without duplication. Schema drift is handled via mergeSchema = true since FHIR resources often evolve. 4️⃣ How do you manage schema evolution across Bronze–Silver–Gold layers? 🏗️ Answer: “In our Medallion Architecture, we validate schema consistency at each layer. Bronze ingests raw JSON, Silver applies type casting + deduplication, and Gold provides curated analytics. To manage evolving schemas, I version schemas in Delta logs and maintain a schema registry ensuring backward compatibility.” 5️⃣ How do you build a unified pipeline for batch claims data and real-time clinical events? 🚀 Answer: “I designed a hybrid ingestion framework in Databricks that handles both historical (batch) data and real-time (streaming) HL7 feeds. Batch loads initialize the Delta tables, while streaming updates are applied using foreachBatch. The architecture supports both CDC and append-only streams for lab results, vitals, and patient admissions.” These real-world approaches have helped me build secure, compliant, and high-performing healthcare data platforms using Databricks, Azure, and Delta Lake. #Databricks #HealthcareDataEngineering #Azure #DeltaLake #MedallionArchitecture #CDC #HL7 #FHIR #DataEngineering #InterviewPreparation Kforce Inc IntraEdge Cardinal Health UnitedHealth Group Lakshya Technologies

10

How I Made Decisions at Amazon In 5.5 years at Amazon, I made thousands of decisions. Some were small. Scope tweaks. Deadline shifts. Meeting structure changes. Some were big. Launch calls. Escalations. Resource tradeoffs. Risk acceptance. And here’s what I learned: Indecision is usually more expensive than a wrong decision. Amazon forced me to build a decision framework…not rely on instinct alone. Here’s how I approached it: 1/ I asked: Is this a one-way door or a two-way door? ↳ A one-way door decision is hard to reverse ↳ Long-term architectural changes ↳ Big resourcing bets ↳ Major customer commitments These required deep dive, multiple perspectives, and slower motion. ↳ A two-way door decision is reversible ↳ Pilot launches ↳ Process changes ↳ Messaging experiments These required speed. Most people treat two-way doors like one-way doors. That’s how velocity dies. 2/ I defined the real risk before acting ↳ What’s the worst-case outcome? ↳ How recoverable is it? ↳ What’s the cost of waiting? Often the cost of delay was greater than the cost of being imperfect 3/ I gathered just enough signal…not perfect clarity ↳ 70% confidence was often enough ↳ The remaining 30% came from moving ↳ Over-analysis was a hidden tax on momentum 4/ I wrote my thinking down ↳ Framing the decision in writing exposed weak assumptions ↳ It also made it easier for others to pressure test it ↳ Writing sharpened judgment 5/ I owned the outcome…even when it missed ↳ If a decision didn’t land, I didn’t blame the inputs ↳ I asked: What did I miss? What pattern did I ignore? ↳ Decisions compound…so does accountability The biggest shift Amazon gave me was this: Decisiveness is a leadership trait. But structured decisiveness is a competitive advantage. 📬 I write weekly about clarity, ownership, and high-stakes execution in The Weekly Sync: 👉 https://lnkd.in/e6qAwEFc What’s one decision you’ve been overthinking that might actually be a two-way door?

34

1 Comment

Post 1: Navigating a Layoff on a Visa? Here Are Some Tips Getting laid off is tough but being laid off while on a visa like OPT, STEM OPT, or H-1B adds another layer of stress. If this is you or someone you know, here are some actionable tips to move forward: 1. Understand your grace period: • F-1 OPT/STEM OPT: 30 to 90 days to find a new job, enroll in a program, or change status. • H-1B: 60-day grace period or until your I-94 expires (whichever is earlier) to transfer or switch status. 2. Update your materials: • Refresh your resume, LinkedIn, and portfolio. • Use strong, quantifiable results and relevant keywords. 3. Start networking immediately: • Let your community know you’re open. • Reach out to alumni, former coworkers, LinkedIn connections, and employee resource groups (ERGs). 4. Be clear and specific when asking for referrals: • Share your resume, the job ID, and a short message they can forward. • The easier you make it for someone to help you, the more likely they will. This process can be overwhelming but take it one day at a time. There is a way forward.

26

Sciences Back in 2019, Amazon India used to hire a lot of brilliant scientists. Back then I was jealous of his privilege. I was arguing with my teammate that “No matter how much research you do or you’d like to do, you’d have to eventually push yourself to complete the task at hand. “ I made myself a bit too explicitly vocal, but well… a kid does stupid things when made to feel cornered. He chose a different path. He chose to make fun of my level in subtle ways and went back to his fictional work. I never saw him again in Amazon. He left within the year. I never read his name in any news. Never read any arXiv paper with his name. There are a lot of things you could learn from this story. One could be that just because you started with a bad set of cards, doesn’t mean you’ll keep holding them forever. Another could be, some people conjure up the will to do what’s necessary to climb the ladder; others just keep changing ladders. But the fact remains… That model is still pretty much the same 12 year old kid it was in 2024. The only thing you don’t realise, is that now it’s on cocaine.

4

True fact The biggest ops wins in data engineering are invisible. No outage. No incident call. No escalation. Pipelines keep running. Alerts stay quiet. Costs stay predictable. And demand is met because capacity was planned before it was needed. Most people notice ops only when something breaks. But real ops impact happens much earlier, every single day. If nothing happened in production this week, someone did a lot of things right. Curious how other teams handle this. What ops practice helped you manage scale or reduce incidents? #dataengineering

8

Positive thoughts to my ex-colleagues at Amazon. A tough period indeed. The first few days after a layoff feel surreal. The silence after the news can be deafening, and the urge to immediately fix everything is strong. But here’s the truth most people don’t tell you , it’s okay to pause. Rest. Let your brain and body catch up to what just happened. Clarity only comes after you exhale. Once you’ve had time to breathe, start treating the next chapter like a strategic reboot. Use this moment to re-evaluate what kind of work excites you and what kind you never want to return to. The layoff might have closed a door, but it also cleared space for something better aligned with who you are now. When you’re ready to re-engage, AI tools can take a lot of the heavy lifting out of the job search. ChatGPT can help you tailor your résumé for each role and practice interview questions. Tools like Rezi and Teal can optimize your resume for applicant tracking systems. Even LinkedIn’s own AI features can help you write stronger “About” sections or suggest better keywords to match your profile with recruiters’ searches. Use this time not just to find another job, but to redesign your career story. A layoff isn’t the end of your professional identity. It’s a forced pause that can turn into your most creative rewrite yet. #Layoffs #CareerGrowth #AItools #JobSearch #Resilience #LinkedInCommunity

10

2 Comments