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.
352 posts
Most of the world's data isn't rows and columns. It's relationships.
Social networks. Molecular structures. Supply chains. Financial transactions. Knowledge graphs. All fundamentally graph-structured. And yet, when I mention Graph Neural Networks at…
By Pranay Mahendrakar Read postThe most expensive mistake in AI isn't a bad model. It's solving the wrong problem.
I've watched teams spend six months building a sophisticated deep learning pipeline when a simple rules-based system would have worked fine. I've also seen the reverse — engineers…
By Pranay Mahendrakar Read postThe first time I generated an image with Stable Diffusion, I sat staring at my screen for ten…
Not because the image was perfect — it wasn't. But because the implication hit me: visual content creation just became democratized. Forever. Diffusion models are conceptually beautiful…
By Pranay Mahendrakar Read postI used to be terrified of ML research papers.
Then someone taught me the 3-pass method, and everything changed. Pass 1 (5 minutes): Read the title, abstract, and look at the figures. Just decide if this paper is relevant to what…
By Pranay Mahendrakar Read postI once deployed a model with 99.2% accuracy. The client was thrilled.
Then we discovered it was a fraud detection model where only 0.5% of transactions were fraudulent. The model was predicting "not fraud" for everything and achieving 99.5% accuracy by…
By Pranay Mahendrakar Read postA common mistake I see engineers make: trying to learn AWS, GCP, AND Azure simultaneously.
You end up knowing three platforms poorly instead of one platform well. My recommendation: pick one, go deep, and be functional in the others. If you're just starting with ML in the…
By Pranay Mahendrakar Read postA healthcare client needed to train a model but couldn't share patient data due to privacy…
Same solution for all three: synthetic data. The ability to generate realistic but fake data is quietly becoming one of the most important capabilities in ML. It solves privacy…
By Pranay Mahendrakar Read postPeople think my 1900+ LeetCode problems are about interview prep. They're wrong.
Well, partly wrong. But the real value of competitive programming for an ML engineer is how it rewires your brain. After a few hundred problems, you start instinctively thinking about…
By Pranay Mahendrakar Read postA client wanted to run a 70B parameter model. Their budget: one consumer GPU.
Three years ago, I'd have said "impossible." Last month, I delivered it. Quantization has changed the game completely. INT8 gets you most of the quality at half the memory. INT4 is…
By Pranay Mahendrakar Read postThe finance industry has a dirty secret: some of the most profitable ML applications are also…
Not algorithmic trading with RL (that's the sexy one everyone talks about). I mean credit risk scoring with XGBoost. Automated document processing for compliance. Transaction monitoring…
By Pranay Mahendrakar Read postI maintain 176+ GitHub repositories.
Code versioned? Yes. Data versioned? No. Model weights tracked? Sometimes. Experiment configs? Scattered across Slack messages and sticky notes. Reproducibility? "Let me try to remember…
By Pranay Mahendrakar Read postWriting 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 post