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
Every model I deploy now gets a model card. No exceptions.
A model card is a one-page document that answers: What does this model do? What data was it trained on? What are its limitations? How well does it perform across different groups? When…
By Pranay Mahendrakar Read postThe number one question I get from junior engineers: "How do you keep up with AI? It moves so fast."
Honest answer: I don't keep up with everything. Nobody does. And accepting that is the first step to staying sane. My information diet: arxiv papers — I scan titles daily, read abstracts…
By Pranay Mahendrakar Read postThe first thing I do on every AI consulting engagement surprises most clients: I try to talk…
Not because I don't want the work. Because starting with "do you actually need AI?" prevents the most expensive mistake in the industry — building a complex ML system when a rules…
By Pranay Mahendrakar Read postI rebuilt a company's internal search from keyword-based to semantic.
Keyword search: user types "revenue growth Q3" — finds documents containing those exact words. Misses the document titled "Third Quarter Financial Performance" that has exactly the…
By Pranay Mahendrakar Read postWorking with ISRO taught me what "production-grade" really means.
In most companies, if your model fails, someone sees the wrong recommendation or gets a bad search result. Annoying, but recoverable. At ISRO, if your system fails, the consequences can…
By Pranay Mahendrakar Read postI built myself an AI second brain. It's the most useful thing I've ever made.
Every paper I read, every article I bookmark, every project note, every technical learning — indexed, embedded, and searchable through a RAG interface. When I'm designing a new system…
By Pranay Mahendrakar Read postThe hardest ML projects aren't the ones with difficult data or complex models.
"Build us an AI that makes our customer service better." Better how? Faster responses? More accurate answers? Higher customer satisfaction? Fewer escalations to humans? Lower cost per…
By Pranay Mahendrakar Read postThe quality of your fine-tuning dataset matters more than the quantity.
We had 50,000 examples for fine-tuning. The model trained fine. Metrics looked decent. But the outputs were inconsistent and sometimes bizarre. The problem: our 50,000 examples were…
By Pranay Mahendrakar Read postWe launched an AI feature to 100,000 users. Here's what went wrong and what we'd do differently.
What went wrong: The LLM cost spiked 4x our projection because users sent much longer inputs than our test cases. Some users discovered they could make the system generate unlimited…
By Pranay Mahendrakar Read postThe most valuable AI engineers I've worked with weren't the strongest coders.
An ML engineer who also understands: database design can build more efficient feature stores. Frontend development can prototype AI-powered interfaces. Product management can prioritize…
By Pranay Mahendrakar Read postThe most dangerous model isn't the one that's wrong. It's the one that's wrong AND confident.
I've seen models output predictions with 99% confidence that were completely incorrect. Users trust confident predictions. Systems downstream act on confident predictions. Confident…
By Pranay Mahendrakar Read postI inherited an ML project with no documentation.
It took me 3 weeks just to understand what the system did. Another 2 weeks to figure out how to retrain the model. Another week to discover there was a critical data preprocessing step…
By Pranay Mahendrakar Read post