Notes from the
edge of the model.
Field notes on what actually breaks in production — agents, retrieval, evaluation, MLOps, and the career decisions nobody writes down. Longer arguments become papers; these are the rest.
353 posts
Writing my third AI book taught me something I didn't expect: teaching AI is significantly…
When you build a system, you can rely on intuition. You make choices that "feel right" based on experience. When you write about it, you have to articulate WHY every choice was made. And…
By Pranay Mahendrakar Read postWe caught a production model silently performing 15% worse for users over 55 last year.
Nobody intended it. The training data just had fewer samples from older users. The model optimized for overall accuracy and effectively deprioritized the underrepresented group. This is…
By Pranay Mahendrakar Read postA client once asked me to build an image classifier for their specific product line.
A few years ago, I'd have said "we need at least 50,000 images." Instead, I took a pre-trained ResNet, fine-tuned the last few layers on their 500 images, and had a working classifier in…
By Pranay Mahendrakar Read postPeople ask me how I stay current in a field that moves this fast.
Morning (before "real work"): one ML paper from arxiv or Hugging Face daily papers — usually just the abstract and figures, maybe the methods section if it's relevant. Two LeetCode…
By Pranay Mahendrakar Read postI once improved a model's performance by 12% without changing the model architecture…
All I did was add three domain-specific features that a subject matter expert suggested over coffee. Feature engineering is the unglamorous skill that wins competitions, ships products…
By Pranay Mahendrakar Read postSelf-driving cars are the most complex ML system that exists. Full stop.
3D object detection fusing LiDAR and camera data. Path planning combining reinforcement learning with graph algorithms. Behavior prediction using transformers. Sensor fusion across…
By Pranay Mahendrakar Read postThe worst bug I ever encountered: a model that ran perfectly, passed all tests, and gave…
No error messages. No crashes. Just confident, consistent, wrong outputs. Took me three days to find it. The preprocessing pipeline applied normalization differently during training and…
By Pranay Mahendrakar Read postStandard RAG has a dirty secret: it treats your documents as isolated chunks of text.
That works fine until someone asks: "Which team members worked on both Project Alpha and the Q3 budget revision?" Now your RAG system needs to understand relationships, not just…
By Pranay Mahendrakar Read postJunior engineers always ask me: "What tools should I learn for MLOps?"
And I always give the same frustrating answer: "It depends." But let me at least give you the framework. Every production ML system needs something from each of these categories…
By Pranay Mahendrakar Read postMy non-technical CEO asked me: "Can I just type 'show me revenue by region for Q4' and get the…
I built it in a week. He uses it every single day now. Text-to-SQL might be the highest ROI AI application for most companies. Every organization has databases. Most business users can't…
By Pranay Mahendrakar Read postI have 240+ certifications. Want to know how many got me a job? Maybe three.
Don't get me wrong — certifications have value. AWS ML Specialty tells employers you understand the platform. Google Cloud Professional ML Engineer is respected. NVIDIA Deep Learning…
By Pranay Mahendrakar Read postA furniture company came to me wanting product photos in 50 different room settings.
What we built: a fine-tuned Stable Diffusion pipeline that generates photorealistic product images in any setting. Cost: about $3,000 in development and a few dollars per batch of…
By Pranay Mahendrakar Read post