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
I failed my first MLOps interview because I could explain model drift conceptually but couldn't…
Theory vs. practice. That gap is exactly what MLOps interviews test. The questions that come up repeatedly: How do you detect model drift in production? (Hint: statistical tests on…
By Pranay Mahendrakar Read postRecommendation systems reportedly drive 35% of Amazon's revenue.
Yet most ML engineers have never built one from scratch. The architecture has evolved dramatically. The modern approach uses a two-tower model: one tower embeds users, another embeds…
By Pranay Mahendrakar Read postControversial take: most production systems don't need GPT-4 class models.
I replaced a GPT-4 API call with a fine-tuned 3B parameter model for a client's customer categorization task. Same accuracy. Cost went from $400/day to $12/day. Latency dropped by 80%…
By Pranay Mahendrakar Read postI spent two weeks tuning hyperparameters on a client project. Improved performance by 0.3%.
Then I spent two days fixing label errors in the training data. Performance jumped 4.2%. Andrew Ng has been preaching data-centric AI for years. Having built 250+ systems, I can confirm…
By Pranay Mahendrakar Read postWorking at ISRO, I see firsthand something that most global tech discourse misses: India isn't…
The opportunities here are massive and unique. Indic language NLP — building models that work in Hindi, Tamil, Marathi, Bengali, and dozens of other languages with hundreds of millions…
By Pranay Mahendrakar Read postA model I deployed for a client started giving increasingly weird predictions.
By the time they flagged it, the model had been confidently making bad decisions for 10,000+ transactions. The input data had shifted gradually — a supplier changed their data format…
By Pranay Mahendrakar Read postHonest take on Quantum Machine Learning in 2026: it's mostly still research. And that's okay.
There are specific areas where quantum advantages are real — molecular simulation, certain optimization problems, quantum-enhanced feature spaces. But for most practical ML tasks today…
By Pranay Mahendrakar Read postWe shipped an LLM-powered feature last year with no evaluation framework.
Within two weeks, users found that it hallucinated medical advice, gave different answers to the same question depending on phrasing, and sometimes just... made up citations. Never…
By Pranay Mahendrakar Read postI reviewed 500 AI job postings last month.
What's disappearing: "Data Scientist" roles that are purely notebook-based analysis. Pure prompt engineering positions. Generic "AI/ML Engineer" roles with vague descriptions. What's…
By Pranay Mahendrakar Read postBy 2028, I believe there will be an "agent economy" — a marketplace where AI agents hire other…
Sound crazy? It's already happening in prototype form. An orchestrator agent receives a complex request. It breaks it into subtasks. It identifies which specialized agents can handle…
By Pranay Mahendrakar Read postAn AI Safety Engineer with 3 years of experience just got offered $280K in the US.
The salary landscape in AI has fractured into tiers that barely resemble each other: Tier 1 — Frontier Model Companies (OpenAI, Anthropic, DeepMind, Google Brain): $250K-$500K+ total…
By Pranay Mahendrakar Read postThe next frontier after LLMs isn't bigger language models. It's world models.
A world model doesn't just predict the next token. It understands cause and effect. It can simulate what happens when you take an action. It reasons about physics, time, and…
By Pranay Mahendrakar Read post