Sourabh Rodagi

Sourabh RodagiOct 28, 2026

Yesterday I read Uber Freight Engineering team's blog on scaling data systems with clear ownership. Here are 3 things stuck with me. (materialised for Data Analysts/Engineers) 𝗖𝗼𝗻𝘁𝗲𝘅𝘁: Uber Freight operates as a massive logistics platform managing over $18 billion in freight, where 𝒓𝒆𝒂𝒍-𝒕𝒊𝒎𝒆 𝒗𝒊𝒔𝒊𝒃𝒊𝒍𝒊𝒕𝒚 across carriers, shippers, and internal operations is critical. To support this, their architecture runs on over 40 highly heterogeneous microservices pulling from nearly 100 𝒅𝒊𝒔𝒑𝒂𝒓𝒂𝒕𝒆 𝒅𝒂𝒕𝒂𝒃𝒂𝒔𝒆𝒔 and external APIs. Their central search platform processes thousands of complex, ad-hoc queries per second across millions of active freight shipments with 𝒑99 𝒍𝒂𝒕𝒆𝒏𝒄𝒊𝒆𝒔 under a few hundred milliseconds and data freshness within seconds. Here are my 3 key takeaways. (materialised for Data Analysts/Engineers) Grain-First Modeling: Define the row-level primary key (like load_id) before writing SQL. Establishing exact granularity upfront prevents fan-out bugs, duplicate counts, and bad joins down the line. Upstream Data Validation: Don't rely on users to spot broken dashboard metrics. Build automated assertion tests directly into ingestion pipelines to catch source schema drift before it hits your reports. Pre-Denormalize for Speed: Stop running heavy multi-table JOINs at query time. Pre-stitching multi-source fragments into flat materialized tables trades cheap storage for sub-second query performance. #uber #dataanalysis #Learning

Sourabh RodagiOct 16, 2026

Everyone wants AI to save them months(time). Very few teams actually get there. Capital One did: 9 months of analysis → 10 days. But AI wasn’t the first step. Clean data was. 𝗧𝗵𝗲 𝗽𝗿𝗼𝗯𝗹𝗲𝗺 Big companies rent thousands of things in the cloud: servers, storage, IP addresses. Projects end. People move on. The bills don’t. Capital One had ~350 kinds of cloud services to check for waste. The catch: “unused” looks different for every kind. Get it wrong, and you switch off something people rely on. By hand, the estimate was 6–9 months. 𝗪𝗵𝗮𝘁 𝘁𝗵𝗲𝘆 𝗱𝗶𝗱 🟩 1. Clean, trusted data first One documented daily list of everything they rent. The AI can only be as good as the data it reads. 🟪 2. AI checks every kind It wrote a rule to spot “unused” and rated its own confidence: high, medium or low. 🟧 3. An expert checks every unsure answer The AI was fast but not always right. It referenced data that didn’t exist and was overconfident. The expert caught it before anything went live. 𝗧𝘄𝗼 𝗰𝗮𝘀𝗲𝘀 𝘁𝗵𝗮𝘁 𝘀𝗵𝗼𝘄 𝘄𝗵𝘆 𝗯𝗼𝘁𝗵 𝗺𝗮𝘁𝘁𝗲𝗿 A switched-off server → AI: unused (high confidence) → expert confirms → savings. A storage bucket untouched for 90 days → AI: unused? (medium) → expert finds it’s still used daily → kept. Case A is where AI saves time. Case B is where the expert saves you from an outage. 𝗧𝗵𝗲 𝗻𝘂𝗺𝗯𝗲𝗿𝘀 → 18–27x faster, even including human review → 12 of 350 kinds (3.6%) held 30–40% of the savings → “Untouched for X days” rules were wrong 40–60% of the time for services that stay in use without changes 𝗠𝘆 𝘁𝗮𝗸𝗲 AI on messy data doesn’t save time. It just makes mistakes faster. Most teams start with the model. The ones who win start with the data. 𝟰 𝘁𝗵𝗶𝗻𝗴𝘀 𝘆𝗼𝘂 𝗰𝗮𝗻 𝘂𝘀𝗲 𝘁𝗼𝗺𝗼𝗿𝗿𝗼𝘄 1. Pick one trusted, documented data source before you add AI 2. Make the AI rate its confidence and explain why 3. Route every medium and low answer to a human expert 4. Trust clear signals (“switched off”) over age-based guesses (“untouched 90 days”) This works far beyond cloud costs: data quality checks, access reviews, policy audits. Agree, or is “data first” overrated? Full Capital One article in the comments. #AI #DataQuality #Analytics #DataEngineering

Sourabh RodagiOct 15, 2026

If I had to relearn SQL from scratch, I’d follow 6 steps. Foundation first. Then real problems. Then building. 1️⃣ Learn the concepts: CS50 SQL (Harvard, free) 2️⃣ Write your first queries: SQLBolt 3️⃣ Drill the core patterns: LeetCode SQL 50 4️⃣ Solve real interview questions: DataLemur or StrataScratch 5️⃣ Practice explaining out loud: Aced peer mock interviews 6️⃣ Build a project. Mine: a clinic marketing research dashboard. → Pulled clinic data from the Google and Yelp APIs → Ran data quality checks in Python + SQL → Stored it in a SQL database → Built the dashboard on top Courses teach syntax. Projects prove you can use it. All links, plus my GitHub, are in the first comment 👇 Which step are you on? #SQL #DataAnalytics #LearnSQL