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
We deployed a hiring model last year that looked great on paper.
Then someone ran it through a fairness audit. It was systematically rating female candidates 12% lower for engineering roles. Not because we told it to — because the training data…
By Pranay Mahendrakar Read postI re-read "Attention Is All You Need" last week. For maybe the tenth time.
Every time I read it, I notice something new. This time it was how elegantly the positional encoding works — a solution that seems obvious in hindsight but was genuinely creative in…
By Pranay Mahendrakar Read postI remember when running a decent language model required renting a cluster of A100s.
Last Tuesday, I ran Llama 3.2 on my laptop. While on a plane. With no internet. The open source AI ecosystem has moved so fast that I think most people haven't fully grasped what's…
By Pranay Mahendrakar Read postEvery time I see a LinkedIn post about some fancy new model architecture, I think about the…
Three weeks. Not building models. Not fine-tuning. Just figuring out why 30% of the date fields were in five different formats and why someone had entered "yes" in a numerical column 847…
By Pranay Mahendrakar Read postI fine-tuned a Llama model on my company's internal documentation last month.
Two years ago, that would've cost thousands in compute. Now it costs less than a nice dinner. QLoRA changed everything. You're training 0.1% of the model's parameters while keeping 99%…
By Pranay Mahendrakar Read postA doctor I work with told me: "Your model caught a tumor that I missed on the first read."
That sentence haunts me — in the best possible way. AI in healthcare isn't a futuristic concept. It's happening now. Cancer detection from medical images that rivals (and sometimes…
By Pranay Mahendrakar Read postA year ago, most ML engineers I talked to had never heard of vector databases.
Here's the quick version of why they matter: modern AI represents everything as high-dimensional vectors (embeddings). To find similar items — whether it's documents for RAG, products…
By Pranay Mahendrakar Read postI saved a client $95/day on inference costs.
Most ML engineers treat PyTorch as a black box. Tensors go in, predictions come out, and whatever happens on the GPU is someone else's problem. That's fine until you're paying $100/day…
By Pranay Mahendrakar Read postLast weekend I built an agent that researches a topic, writes a draft blog post, critiques its…
A year ago, that was a research project. Now it's a weekend build. The agent framework landscape has matured incredibly fast. LangGraph for complex stateful workflows. CrewAI for…
By Pranay Mahendrakar Read postUnpopular opinion: "Prompt Engineer" as a job title has maybe 18 months left.
Not because prompting doesn't matter — it absolutely does. But because it's being absorbed into everything else. When every developer uses structured outputs, function calling, and tools…
By Pranay Mahendrakar Read postI avoided learning Kubernetes for two years.
Then I needed to deploy a model that scaled from 100 to 10,000 requests per hour depending on the time of day. And I realized I was the bottleneck on my own team. Here's the reality…
By Pranay Mahendrakar Read postI built wildlife detection systems for railway tracks at ISRO.
That project changed how I think about AI. Most AI discourse is about chatbots, image generators, and productivity tools. But some of the most important applications are in climate and…
By Pranay Mahendrakar Read post