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
Neural networks are "inspired by the brain" the way airplanes are "inspired by birds."
They share the fundamental idea (learning from data / generating lift) but the implementation is completely different. And understanding WHY they work requires math, not biology. If you…
By Pranay Mahendrakar Read postThe scariest production bug I ever encountered: the model was right 99% of the time and…
It was a medical triage system. Worked beautifully on routine cases. But for rare conditions — the ones where getting it right matters most — it confidently misclassified them as…
By Pranay Mahendrakar Read postI consulted for an AI startup that raised $5M, hired 12 engineers, and spent 8 months building a…
They ran out of money before launching. A competitor with 3 engineers fine-tuned Llama, built a solid product layer on top, and captured the market in 4 months. The lesson I keep…
By Pranay Mahendrakar Read postNobody talks about the most expensive part of machine learning: labeling data.
I once spent $40,000 on data labeling for a single project. And that was cheap compared to what some teams spend. Here's what I've learned about making labeling sustainable: Start with…
By Pranay Mahendrakar Read postI burned out twice in my AI career. Both times, I saw it coming and ignored the signs.
First time: working 14-hour days for three months straight on a defense project. I told myself it was temporary. It wasn't. My code quality dropped. My decision-making got worse. I…
By Pranay Mahendrakar Read postMost ML engineers build great models and terrible APIs.
I've consumed ML APIs that return predictions in five different formats depending on input. That require 30 lines of preprocessing before you can send a request. That return cryptic…
By Pranay Mahendrakar Read postI pair program with AI now. Not as a gimmick — as my actual daily workflow.
GitHub Copilot writes the boilerplate. Claude helps me think through architecture decisions. I use AI to generate test cases I wouldn't have thought of. And I write the core logic…
By Pranay Mahendrakar Read postInherited an ML codebase last year.
The previous engineer had left. Nobody else understood the system. The model was running in production, making real decisions, and nobody could explain how it worked or reproduce the…
By Pranay Mahendrakar Read postA client's chatbot was giving garbage responses to anything written in Hindi.
Here's what most engineers miss: the tokenizer is the foundation of every language model, and it's also the most common source of subtle bugs. English text gets tokenized efficiently…
By Pranay Mahendrakar Read postI cut a client's AI inference costs from $3,200/month to $180/month.
Step 1: Replaced GPT-4 calls with a fine-tuned Llama 7B for their specific use case. The task was structured extraction — overkill for a frontier model. Step 2: Added aggressive caching…
By Pranay Mahendrakar Read postThe best technical decision I ever made was writing a design doc BEFORE writing code.
It was a RAG system for a financial services client. Complex requirements. Multiple data sources. Strict compliance needs. My instinct was to start coding immediately. Instead, I spent…
By Pranay Mahendrakar Read postI once built a semantic search system where "dog food" returned results about "hot dogs." The…
Embedding models capture similarity in ways that don't always align with what humans mean by "similar." This is the gap that trips up most engineers building search and RAG systems…
By Pranay Mahendrakar Read post