Notes
Random updates and thoughts: what I am building, learning, and thinking about. For longer, specific blogs, read me on Explaining Agents and Medium.
- AI agent memory, part 2: what happens when facts change
Five more experiments on my personal agent's memory. Overwriting old facts left it unable to notice changes, no memory design caught my outdated questions until one prompt sentence took that from 3% to 91%, and a fix for invented projects took them from 3 to 0.
- AI agent memory: 5 research based experiments on an agent with real data
Five experiments on two months of my real coding-session history, testing ideas from the Mem0 and MemoryTax papers on a free-tier model. Overwriting memory destroyed history; facts plus raw text plus mixed keyword and embedding search did best.
- Building a personal AI agent that keeps track of everything I work on
Saati is a personal AI agent I talk to on Telegram. It tracks my projects, ideas, and decisions, follows my coding sessions and voice notes, and runs in the cloud on a Gemini model.
- How gradient descent works in AI: loss, slopes, and the chain rule
A model learns by measuring how wrong it is and nudging every number towards less wrong. This note builds that from nothing: what a loss function is, what a slope really measures, why the gradient points uphill, and how the chain rule prices a weight buried ten layers deep.
- Building Aditi, a Hindi and English voice agent with Pipecat and tool calling
A real-time voice agent for a fictional Ahmedabad real estate agency. She searches listings, works out EMIs, books site visits, and texts a follow-up, in Hindi or English, on a laptop CPU and free tiers.
- How a neural network layer works: matrices, bias, and ReLU
A neural network layer is a matrix, a bias, and a nonlinearity. This note builds it from nothing: what a matrix does to a vector, why stacking linear layers collapses into a single layer, and why ReLU is what makes depth worth having.
- What are word embeddings? How text becomes vectors
A computer has no notion of meaning or context. This note builds word embeddings from nothing: why one number per word fails, why vectors work, and why the angle between two vectors tells you a cat is closer to a dog than to a car.
- Building AI Stack Advisor, a build-vs-buy calculator for AI
Should a business buy a SaaS tool, build on APIs, or run its own model? A free browser tool that answers with costs, TCO, break-even, and ROI, in your own currency and with your country's tax rules.
- Building TickerBrief, an automated equity research pipeline with n8n, Python, and a local LLM
Type a ticker, get a full equity research report with charts, filings, news sentiment, and an LLM-written narrative. How the data, the scorecard, and the three LLM passes work, and why the LLM never touches a number.
- Building a model drift monitor with Evidently, n8n, and a local LLM
A local MLOps stack that notices when a model starts going wrong and explains why, in plain English, on Slack. The metrics, the alert rules, four failure scenarios, and what the test runs showed.
- Building a fully local RAG document assistant with n8n, Ollama, and ChromaDB
A document Q&A assistant that runs entirely on my machine. How it splits, embeds, and retrieves text, the prompt that stops a small model from guessing, and what would change with a cloud model, different chunk sizes, or conversation memory.
- FocusPulse: a Chrome extension for tracking your own focus
A Chrome extension where you track your own focus in four colours. Ideas to make it more addictive are welcome.
- How AI benchmarks are actually designed, from first principles
Benchmarks look like simple tables of percentages. Underneath, each one is a measuring instrument with design choices about tasks, scoring, statistics, and leakage. Here is how they work.
- EasyAutoFill: a browser extension that fills any form from your own profile file
A local, open-source browser extension that fills any form from your own profile file.
- Starting a statistics series on Medium
Stats_with_FirstPrinciples: rebuilding statistical thinking from the ground up, one intuitive post at a time.
- Automating my morning arXiv routine with n8n, Telegram, and an LLM
Checking arXiv every day and feeding papers to an LLM ate my mornings, so I built an n8n workflow to do it. How it works, the bugs that made it silently return nothing, and how it went from a small local model to full-paper derivations.
- How machine learning and deep learning are different
Not "deep learning is ML with neural networks," but why. A first-principles look at what each one is really doing, and why the difference comes down to who builds the representation.
- How AI labs actually train their models
From raw web text to a helpful assistant: the stages every big lab goes through, and how OpenAI, Anthropic, Google, Meta, and DeepSeek each do it differently.
- Why Karpathy said gradient descent writes better code than you
In 2017 Andrej Karpathy tweeted that gradient descent can write code better than you, then made the case in his Software 2.0 essay. What he actually meant, what the essay gets right, what it admitted would go wrong, and how the claim reads nine years later.
- How Claude overtook GPT (where it matters)
ChatGPT is still the household name, but in coding and enterprise use, Anthropic's Claude has caught up and pulled ahead this year. Here is how that happened.
- The maths and physics hiding inside LLMs
Like many others, I was surprised the physics Nobel went to Geoffrey Hinton. So I went looking for the physics and maths inside large language models.
- What is Google's NotebookLM, and how is it different from a chatbot and from RAG?
NotebookLM, ChatGPT, and RAG get discussed as three competing things. They are not. One is a product, one is a pattern, and the difference that actually matters is who decides what the model is allowed to read.
- What is RAG (retrieval augmented generation), and why does it matter so much?
Retrieval-augmented generation lets an AI answer from your own documents instead of its memory. Here is how it works, why it has become so important, and where it goes wrong.
- My quickest paper, and an unexpected collaboration
How a comment in a research meeting turned into a paper with Mustapha Azreg-Aïnou in about six weeks.
- GPT: the model that started it all
With GPT-4 out, I went back to where the GPT series began. The surprising part is how simple the core idea is, and how far scaling it went.
- ChatGPT: the interface that changed everything
ChatGPT is two weeks old and everyone around me is talking to it. The model behind it is not brand new. The way we get to use it is.
- Where machine learning meets black hole energy extraction
Machine learning doesn't calculate the Penrose process for us, but it is quietly becoming essential to everything around it.
- My first paper: stealing energy from spinning black holes (and their mimickers)
The Penrose process in plain words, and what my first paper found when we ran it around Kerr black hole mimickers. One family matched Kerr exactly; another was far more efficient.
- My MSc thesis: following light and matter around black holes
Five months of learning how light and particles move around black holes, naked singularities, and wormholes.