Kirti Pande

H1B Visa fees 100,000 Dollar. Time to awake indians. No other way then to get out of delusion of development. We should focus more on economy then cultural activities going on now a days. This visa row is indication of following 1. AI is really on the way to destruct and change IT sector forever. 2. Jobs placements will be lower then expected. 3. Even if this visa fees is reduced then also there will be an uncertainty on the AI changing the world order. 4. Narratives build that now india will be great again, is false i think. 5. Only way to avoid heavy losses is to truly invest in AI research.

18

Eval is not unit testing. If you treat evals like static pass/fail scripts, you’re missing their real purpose. Think of evals as a PRD for your AI agent’s behavior: a living, evolving contract for what “good” means in production. And like any PRD, it should have observability built in so you can track outcomes, pinpoint weaknesses, and refine in an automated, continuous loop. Improvement without measurement is limited and imprecise. Evals give you the clarity you need to systematically improve your app, not limited at the feature level, but across the full agent lifecycle. They tell you where you’re drifting from the product’s intent, why performance is slipping, and how to course-correct with precision. Instead of snapshot testing, you’re running a feedback engine that constantly learns from both successes and failures. In the application layer, it’s about setting big-picture goals, like “help a customer dispute a fraudulent transaction” or “recommend the right investment product for their profile.” Then you throw everything at it half-filled forms, conflicting details, urgent time limits and see if it still gets the job done. The point isn’t to check if it spits out a canned line, it’s to see if it can solve the problem accurately, fast, and in a way that stays within compliance and cost boundaries. My recommendation approaches to make evals work for AI agents: 1. Define capabilities, not scripts – Focus on what success looks like rather than exact matches. 2. Simulate real-world variability – Feed in unpredictable, noisy, and edge-case inputs. 3. Instrument for observability – Capture reasoning traces, model calls, and decision trees for explainability. 4. Automate refinement loops – Turn eval failures into direct triggers for retraining, prompt tuning, or guardrail updates. 5. Score across multiple dimensions – Measure accuracy, latency, cost efficiency, compliance, and user experience impact in parallel. When done right, evals validate your AI agents and steer them. They transform from a passive QA step into the heartbeat of your improvement cycle, ensuring every release moves the agent closer to its ideal, measurable definition of “good.”

40

3 Comments

Quantization in Large Language Models (LLMs) As LLMs scale, intelligence is no longer the bottleneck — memory, latency, and cost are. Quantization turns out to be one of the most powerful (and misunderstood) techniques enabling real-world deployment of large models. In this blog, I break down: • Why LLMs are surprisingly quantization-friendly • INT8 vs INT4 vs NF4 — what actually works in practice • How modern architectures absorb low-bit noise • Where quantization fails and how to design around it Quantization is no longer an optimization step — it’s foundational infrastructure for production LLM systems. Link: https://lnkd.in/gqGje75G #LLMs #AIEngineering #Quantization #MachineLearning #GenAI #MLOps

9

2 Comments

This CEO broke down the tech behind 10-minute delivery, and it's a masterclass. Everyone sees the delivery. Nobody sees the engineering - inventory prediction across dark stores, real-time routing, demand forecasting at scale. The Blinkit team built something incredible. Must-watch if you're in tech and want to understand how quick-commerce actually works. Kudos to Vivek Ravisankar for this breakdown.

340

6 Comments

𝗪𝗵𝗮𝘁'𝘀 𝘁𝗵𝗲 𝗱𝗮𝘁𝗮 𝗮𝗻𝗱 𝗔𝗜 𝗺𝗮𝗿𝗸𝗲𝘁 𝗯𝗲𝗲𝗻 𝘂𝗽 𝘁𝗼 𝗶𝗻 𝘁𝗵𝗲 𝗳𝗶𝗿𝘀𝘁 𝟮 𝘄𝗲𝗲𝗸𝘀 𝗼𝗳 𝟮𝟬𝟮𝟲? Here are the emerging trends in the data and AI talent landscape based on my meetings with Heads of AI and CDOs along with a note on the churn at C-level ✅ High demand for hands on data and AI engineers which is also driving higher salary packages ✅ High demand for data scientists that have put GenAI products and solutions into production - this is fueled to a certain extent by AI CoEs ✅ High demand for AI governance specialists ↘️ Low demand for data scientists that have limited experience with putting AI solutions into production ↘️ Low demand for hands off managers ↘️ Low demand for pure data analysts. While the head of AI/ CDO function is still in its infancy in many organizations. Which means, data/ AI maturity is low especially from a culture perspective. However, expectations on what these teams need to deliver are sky high because of the narrative built externally based on how companies want to be perceived by their shareholders/ customers. As a result, the pressure on these teams and their leaders is immense. This often leads to a revolving door situation for CDO/ head of AI roles. The prediction I made late last year regarding change of guard at the CDO level is already playing out in many banks and FS companies. ♻️ If you found the insights valuable, please share it so that people from your network can benefit 📍 Except for the image, no AI was used to create this post!

17

Yesterday I was talking with some folks about what kinds of questions they ask during interviews for AI/ML/DS/Data roles.  There were a bunch of the more obvious technical questions about tradeoffs between different techniques.  We then got to discussing how to understand how a candidate would approach a more realistic problem faced in the normal day to day work.  One of the questions I often like to ask is “How would you improve model reliability if you couldn’t change the model itself?” I like to ask this since it is fairly open ended and requires the candidate to actively think about the question.  It is in line with how managers in real life will often ask for improvements with ambiguous requests. Good starting questions from the candidate would revolve around what does reliability mean?  Precision gains? Latency reductions? Resilience to bad inputs?  The answer depends on the specific context of the business problem. We often then move on to various ways to focus on data improvements. Is the evaluation data good?  Are edge-cases getting missed? You can't touch the model, but you can tighten the feedback loop.  Maybe you can introduce upstream filters to catch problematic inputs (corrupt images, out-of-distribution text, sensor artifacts) all before they reach the model. The next focus area can be to wrap the model in better logic or pre/post processing.  You can use a secondary model or lookup system to override or veto low-quality predictions.  Embed the model in a rules-based framework that ensures consistency with known constraints (e.g., safety boundaries in autonomous driving). Some of the best candidates include ideas like working with downstream teams (e.g., product, ops, annotation) to ensure error cases are being flagged and fed back into future retraining or retriage loops.  Often, problems can get addressed, not just with technical solutions, but instead with better collaboration and partnerships.  Candidates can take these questions in many directions and have given me great insight into how they would address real world problems. I am always curious about the kinds of questions others ask to cover these kinds of topics.

34