Sai Krishna K

With so much conversation around AI capabilities, the real challenge is moving from experimentation to production, responsibly and at scale. Uchenna Mgbaja, PMP® ’s focus on applied AI frameworks and bridging the research-to-operations gap is exactly what the ecosystem needs more of. Also appreciate the emphasis on ethical and equitable AI, especially across public, private, and nonprofit sectors. Responsible deployment isn’t a bolt-on, it has to be built into the framework. If you’re in Houston or joining virtually, this should be a highly practical session. 📅👩‍💻 #WomenInData #AppliedAI #MachineLearning #EthicalAI #DataLeadership

4

Having worked with a wide variety of LLMs, this is a great roadmap to follow for anyone who wants to truly learn NLP

9

🚀 Partitioning vs Clustering in Data Engineering — Made Simple In large datasets, smart data organization can turn slow queries into lightning-fast results. Two common techniques: Partitioning and Clustering. 1️⃣ Partitioning – “Separate folders” We split a table into chunks based on a column value. Example: 📂 /sales/country=USA/... 📂 /sales/country=UK/... Query: SELECT * FROM sales WHERE country = 'USA'; ✅ Only scans the USA folder — skips all UK data. Best for: Columns with few unique values (e.g., Country, Year). 2️⃣ Clustering – “Organized shelves” We keep the data in the same file but group similar rows together. ✅ Without Clustering The data might be stored unsorted, like: file1.parquet → (2024-01-03, USA), (2024-01-01, USA), (2024-01-02, UK)... file2.parquet → (2024-01-02, USA), (2024-01-01, UK), (2024-01-03, UK)... ✅ With Clustering by order_date The data inside each file is grouped and sorted by order_date: file1.parquet → 2024-01-01, USA, 100 2024-01-01, USA, 150 2024-01-02, UK, 200 2024-01-02, USA, 180 file2.parquet → 2024-01-03, UK, 220 2024-01-03, USA, 300 ✅ Queries can skip irrelevant ranges inside the file. Best for: Columns with many unique values (e.g., Order Date, User ID). 3️⃣ Together = Maximum Efficiency In BigQuery or Snowflake, you can do: PARTITION BY DATE(order_date) CLUSTER BY country This way: Partitioning skips entire days. Clustering skips irrelevant countries inside each day. 💡 Rule of Thumb: Partition for broad pruning (skip whole files) Cluster for fine-grained pruning (skip parts of files) 🔍 How do you use partitioning and clustering in your data projects? #DataEngineering #BigQuery #Snowflake #Partitioning #Clustering #SQLPerformance #DataModeling

15

With the rapid rise of GenAI and Agentic AI trainings, it can be challenging to identify programs that truly balance theory with real-world application. Recently, my Data Science team at Target In India and I attended an abridged two-day Agentic AI training led by Abhijith Neerkaje and Ajay Shenoy —both former colleagues whose deep retail and data science expertise significantly elevated the experience. Drawing from their hands-on leadership of retail data science teams, they clearly explained core concepts while grounding them in practical use cases such as supply chain and pricing optimization. The sessions were engaging, practical, and rich with insights directly applicable to live projects. Their clarity, patience, and focus on real-world relevance stood out. I highly recommend their full Agentic and Generative AI programs to anyone looking to effectively bridge theory and industry application.

45

1 Comment

Most of us may have faced one question on Bayesian inference (formula below) in our data science interview where we needed to find out the posterior, given the other 3 metrics. p(theta | D) = p(D | theta) × p(theta) / p(D) , p(D) = ∫ p(D | theta) × p(theta) d theta Simple and straight, right. In a few special cases, it is also computationally complete. For example, with conjugate models (Normal prior + Normal likelihood), the posterior stays Normal, the denominator integral can be solved analytically, and we get closed-form expressions for posterior mean and variance. But this is textbook Bayesian inference. In most real ML problems - logistic likelihoods, nonlinear models, hierarchical structures, latent variables, or high-dimensional parameters - the denominator p(D) becomes intractable. The posterior, which combines prior belief with data evidence still exists, but we cannot normalize it, integrate it, or compute expectations directly. That’s why Bayes formula isn’t the end. That gap is why modern Bayesian inference relies on optimization and sampling methods. Variational Inference (VI): Bayes as Optimization - Instead of computing the true posterior p(theta | D), we choose a simpler distribution q(theta) (often Gaussian) and make it as close as possible to the posterior by minimizing KL divergence: KL( q(theta) || p(theta | D) ) This seems impossible because the posterior contains the intractable denominator. The key insight is rewriting the objective into the Evidence Lower Bound (ELBO): log p(D) = ELBO(q) + KL(q || p) Since KL ≥ 0, maximizing ELBO is equivalent to minimizing KL - without computing p(D). - q(theta) is our approximate posterior - ELBO maximization adjusts its parameters - Expected data fit + prior regularization − uncertainty penalty VI replaces difficult integrals with gradient-based optimization. It is fast and scalable, but accuracy depends on how expressive q(theta) is and often underestimates uncertainty. MCMC: Bayes as Sampling - Instead of approximating the posterior, it samples from it. In Metropolis–Hastings: 1. Propose a new value theta* 2. Compute an acceptance ratio r = ( p(D | theta*) × p(theta*) ) / ( p(D | theta_t) × p(theta_t) ) 3. Accept with probability, alpha = min(1, r) Crucially, the denominator p(D) cancels out. If alpha is high, the move is accepted; if low, it may be rejected. Accepting occasional worse moves prevents the chain from getting stuck and guarantees correct long-run behavior. MCMC is asymptotically exact but computationally heavier and requires convergence diagnostics. - VI: optimization-based, fast, scalable, approximate - MCMC: sampling-based, slower, but theoretically exact Both exist for the same reason: The denominator in Bayes’ formula is usually impossible to compute. And so, Bayesian inference doesn’t stop at Bayes’ rule and the real-world inference begins. #Bayesian #Optimization #Sampling

22

1 Comment

I’m reaching out to companies building in the AI space. ACM Dallas is hosting a technical workshop on Observability & Tracing in AI Agents, and we’re opening sponsorship opportunities. This isn’t a generic AI talk. The audience will be engineers, researchers, and practitioners working on real production systems. If your company works on: AI infrastructure Monitoring platforms Developer tooling LLM frameworks Cloud / MLOps platforms This is a great opportunity to connect with a highly technical community.

12

🚀 𝘿𝙖𝙮 47 𝙤𝙛 100 𝘿𝙖𝙮𝙨 𝙤𝙛 𝘿𝙖𝙩𝙖 𝙎𝙘𝙞𝙚𝙣𝙘𝙚: 𝙒𝙝𝙖𝙩 𝙄𝙨 𝙋𝙧𝙚𝙘𝙞𝙨𝙞𝙤𝙣-𝙍𝙚𝙘𝙖𝙡𝙡 𝙏𝙧𝙖𝙙𝙚-𝙤𝙛𝙛? 🔍 𝙄𝙣𝙩𝙧𝙤𝙙𝙪𝙘𝙩𝙞𝙤𝙣 When dealing with imbalanced classification tasks, accuracy isn’t enough. Understanding the precision-recall trade-off helps you evaluate how well your model identifies positive instances without over-predicting them. Let’s break it down. ⚠️ 𝗣𝗿𝗼𝗯𝗹𝗲𝗺 𝗦𝘁𝗮𝘁𝗲𝗺𝗲𝗻𝘁 In real-world applications like spam detection or medical diagnosis, false positives and false negatives have different costs. How do we choose a threshold that balances precision and recall effectively? 💡 𝗦𝗼𝗹𝘂𝘁𝗶𝗼𝗻 🔹 Precision: Proportion of predicted positives that are correct. 🔹 Recall: Proportion of actual positives that are correctly identified. 𝗔𝘀 𝘆𝗼𝘂 𝗶𝗻𝗰𝗿𝗲𝗮𝘀𝗲 𝗽𝗿𝗲𝗱𝗶𝗰𝘁𝗶𝗼𝗻 𝘁𝗵𝗿𝗲𝘀𝗵𝗼𝗹𝗱: Precision increases, fewer false positives Recall decreases, more false negatives The goal is to find the right balance based on business needs. ✅ 𝗣𝗿𝗼 𝗧𝗶𝗽 Use F1 score to automatically balance precision and recall. For cost-sensitive domains, define a custom threshold based on domain knowledge. 📚 𝗥𝗲𝗳𝗲𝗿𝗲𝗻𝗰𝗲𝘀 & 𝗟𝗶𝗻𝗸𝘀 Scikit-learn Precision-Recall Curve - https://lnkd.in/gwBPxKCH ROC vs PR Curves – Machine Learning Mastery - https://lnkd.in/gnYM3hsZ #Classification #EvaluationMetrics #DataScience #MLInterview #100DaysOfDataScience

7

Building an LLM - concepts and interview question series Topic 5 – Quantization: PTQ vs. QAT When asked, “How do you optimize a 70B parameter model for inference?” many candidates jump to “Quantization.” But very few can explain how we get from FP16 down to INT8 or INT4 without turning the model into gibberish. The bridge between massive weights and limited GPU VRAM is carefully engineered quantization. So to control the trade-off between perplexity (accuracy) and latency, you must distinguish between two paradigms: Post Training Quantization (PTQ) - This is like a quick win. You take a pre-trained Llama-3 or Mistral checkpoint and calibrate it after training. Mechanics: Uses a small dataset to fix activation ranges. Used - Flows like load_in_8bit or GGUF fall here. (for more see bitsandbytes package documentation) Reason it works: Large models have parameter redundancy. They can "eat" the precision loss with <1% perplexity degradation. Quantization-Aware Training (QAT) – So here you inject fake quantization nodes during a fine-tuning stage. The model “sees” the rounding noise during the forward pass and adapts its weights to survive it. (Fake it till you make it for LLMs :)) Used: Critical for INT4 (and below) or on-device SLMs. Reason it works: At 4-bit precision, naive PTQ often breaks reasoning capabilities. QAT recovers that lost signal. Why this matters for LLMs: A 7B model in FP16 needs ~14GB VRAM. In INT8, it needs ~7GB. This isn't just about saving space; it is about Memory Bandwidth. Since inference is memory-bound, moving 50% less data per token directly translates to higher tokens-per-second. Real-World Example: Meta’s DLRM Recommender systems at scale (like Meta’s) are often bottlenecked by fetching massive Embedding Tables from memory, not by the CPU compute. By applying QAT to these embeddings, they compress Terabyte-scale tables by 4x. This frees up memory bandwidth, allowing the model to run drastically faster in production without losing the nuances about user preferences. My Rule of Thumb: Start with a solid PTQ process. If perplexity degrades beyond tolerance (especially at 4-bits)—pay the compute cost for QAT. The code block below shows how PyTorch simulates INT8 noise during training (the QAT workflow). #PyTorch #LLM #GenerativeAI #MachineLearning #Quantization

23

3 Comments