Smriti Mishra

Smriti MishraOct 13, 2026

500 visits to your pricing page sounds promising. But my first question would be: who are those 500 visits actually from? Potential customers? Existing customers? Companies already considering you? Or people just having a look? While exploring Snitcher, one thing stood out to me: two companies can visit the exact same pricing page, but those visits can mean very different things. One could be an existing customer checking something. Another could be a company that fits your ideal customer profile and has been coming back to your website repeatedly. In a typical analytics dashboard, they can both look like just another page view.That’s the part of Snitcher I found useful. It adds company-level context to some of that anonymous traffic, so you can start distinguishing between existing customers, potential prospects, companies showing stronger intent, and accounts becoming more active on your site. For me, that’s when the data becomes more useful because you have more context for deciding what to do next. Sales can use that context to help prioritize accounts. Marketing can get a better sense of whether campaigns are attracting the kinds of companies they actually wanted to reach. Of course, a visit to a pricing page doesn’t automatically mean someone is ready to buy. But knowing a little more about the company behind that visit gives you much more to work with than a page-view count alone. How do you figure out who’s behind your pricing page traffic right now? Genuinely curious.

Smriti MishraOct 2, 2026

The post with the most impressions wasn’t the one with the highest engagement rate. I gave Viktor three LinkedIn analytics exports in Slack and asked him to turn them into a performance report. He returned an Excel workbook with comparisons, charts, and ideas to test next. Here is the task, what came back, and my take! Get started free at viktor.com. $100 in credits included, no card. Full link in first comment. In partnership with Viktor. #artificialIntelligence #technology

Smriti MishraOct 5, 2026

A model can write valid SQL and still calculate the wrong business metric. A revenue query might double-count sales after a join, overstating performance in a report. Accuracy in analytics code generation matters because someone will use those numbers to make a decision, which is where model customization with Amazon SageMaker AI can help. Whether you work in data engineering, data science, or ML engineering, those errors can mean extra checks, manual corrections, and delayed insights. That is what interests me about model customization: making outputs more reliable for the work we actually need them to do. I like to think of it as a simple rule of thumb: consider RLVR when you can verify the result, and DPO when you can show which response better meets your quality criteria. 1/ Reinforcement learning with verifiable rewards (RLVR) RLVR uses verifiable checks as feedback during training. For SQL generation, you could compare query outputs against expected results on controlled datasets, including cases with duplicate records or missing values. Just checking whether the query runs would miss quite a lot. The checks need to establish whether it calculates the right metric using the agreed business rules. 2/ Direct preference optimization (DPO) DPO trains on pairs of preferred and less-preferred responses. Staying with SQL, two responses might return the correct result, but one has a clearer query structure, states its assumptions, and explains the business logic in terms your team understands. Those preferences can become training pairs, helping shape outputs that are easier to review and maintain. The part I find most interesting is the work behind that feedback. Expected results need to be trustworthy, and reviewers need to agree on what makes one response better. Amazon SageMaker AI supports both approaches through serverless model customization, managing the underlying compute so teams can focus on data, training, and evaluation. To evaluate improvement, I would compare the base and customized models on the same unseen SQL tasks. I would measure how often they return the correct results and how much time reviewers spend correcting them, then weigh those improvements against training and inference costs. You can try the Serverless Model Customization with Amazon SageMaker AI workshop, with hands-on labs for DPO and RLVR, alongside deployment and evaluation: https://fandf.co/4cZPHaH Then put those skills into practice through the AWS AI League model customization challenge: https://fandf.co/4hsc61D Thanks to Amazon Web Services (AWS) for collaborating on this post. #ad #artificialintelligence #machinelearning #dataengineering #technology