# Yashovardhan Srivastava > Builder. Researcher. I rely on obsession, not talent. Yash writes about program search with LLMs (cadence), mechanistic interpretability, ML systems, and developer tools he builds. Code: https://github.com/yash-srivastava19 ## Posts - [(Non Technical) On the Ancient Sport of Hacking](https://yash-sri.xyz/blog/sport_hacking): What programming actually is, and why taste matters: an essay on the ancient sport of hacking. - [(Non-Technical) I'm a Tool Guy](https://yash-sri.xyz/blog/tool_guy): On language and tools as what made us human, and why I call myself a tool guy. - [(Non-Technical) On my experience with hiring and making your career identity](https://yash-sri.xyz/blog/career): Lessons from job hunting as a new graduate: why your career identity shouldn't become your whole identity, and how to find work where you thrive. - [(Non-Technical) The Art of Ignoring What Everyone Else Assumes](https://yash-sri.xyz/blog/first_principles): On first-principles thinking: a practical structure for questioning what everyone else assumes, so you save your brain power for decisions that matter. - [(Non-Technical) The Unemployed Developer](https://yash-sri.xyz/blog/unemployed_developer): Why can't developers who build and ship every day find jobs? An attempt to make sense of the 2026 developer job market. - [Arrakis - A toolkit to conduct, track and visualize mechanistic interpretability experiments](https://yash-sri.xyz/blog/arrakis_blog): Arrakis is an open-source toolkit to run, track, and visualize mechanistic interpretability experiments on language models. How it works and why I built it. - [Cadence: Program Evolution via Large Language Models](https://yash-sri.xyz/blog/cadence_blog): Cadence uses LLMs to evolve programs, generating, mutating, and scoring candidates to solve problems like the Traveling Salesman Problem. Design and results. - [Design for Generality](https://yash-sri.xyz/blog/generality): Design is complicated, not complex: a short essay on reducing moving parts without losing generality, with Transformers as the example. - [Do LLMs Have a Secret Interlingua? Testing GPT-2's Hinglish Comprehension with Sparse Autoencoders](https://yash-sri.xyz/blog/sae_macaronic_blog): Does GPT-2 understand Hinglish? Using sparse autoencoders to test whether a language model shares features across Hindi and English words. - [Elixr: What if PyTorch Used Complex Numbers?](https://yash-sri.xyz/blog/elixr_blog): Elixr is a micrograd-style autograd engine over complex numbers. What changes about backprop when values are complex, and what I learned building it. - [Flex Attention - How to Scale Attention Models for Billion Users?](https://yash-sri.xyz/blog/flex_attention): An introduction to PyTorch's FlexAttention: how it makes custom attention variants fast without writing kernels, inspired by Horace He's talk at Jane Street. - [I built Git from scratch. Here's what I learned.](https://yash-sri.xyz/blog/verizon_blog): I built Git from scratch to really understand it. What building Verizon, a Git clone, taught me about objects, refs, and HEAD. - [JAX - Why is everyone excited about this framework?](https://yash-sri.xyz/blog/jax_blog): What makes JAX different from PyTorch and TensorFlow: a from-scratch walkthrough of JAX's core ideas, simpler than the official Autodidax tutorial. - [Leviathan - Let's modify and improve the Transformer model(from scratch)](https://yash-sri.xyz/blog/leviathan): Leviathan replaces the Transformer's attention scores with a correlation score borrowed from signal processing, aiming for a drop-in replacement. Built from scratch. - [On High Agency and Work Ethic](https://yash-sri.xyz/blog/agency): A short essay on high agency and work ethic: why ownership and obsession beat waiting for genius or inspiration. - [REALBT - A Simple Backtesting Engine in Pure Python](https://yash-sri.xyz/blog/realbt): REALBT is a simple backtesting engine for trading strategies, written in pure Python. How it works and what I'd improve. - [SDPA Optimization Case Study - Worklog](https://yash-sri.xyz/blog/sdpa_optim): A worklog optimizing scaled dot-product attention (SDPA) in PyTorch with torch.compile and benchmarks, written to learn PyTorch internals. - [Safe SQL: A Seatbelt for Your Database](https://yash-sri.xyz/blog/safe_sql_blog): Safe SQL catches dangerous queries, like an UPDATE without a WHERE clause, before they reach your database. A Python package and CLI. - [Shoving NEAT into JAX: When Evolving Topologies Meet Static Computation Graphs](https://yash-sri.xyz/blog/neat_jax_blog): Implementing NEAT (NeuroEvolution of Augmenting Topologies) in JAX for EvoJAX: how networks with evolving topologies fit static computation graphs. - [The Harness Was the Experiment: what four benchmark problems taught me about program search](https://yash-sri.xyz/blog/evolving_problems_cadence): A worklog running cadence, an LLM program-search library, on four published benchmarks with 20 model calls a day, and why the harness mattered more than the search. - [Using MCTS to Navigate Sparse Autoencoder Feature Space](https://yash-sri.xyz/blog/deeprobe_blog): Deeprobe uses Monte Carlo Tree Search to explore sparse autoencoder feature space instead of a game board. What it found and how it works. - [What Happens When You Run BPE on Encrypted Text?](https://yash-sri.xyz/blog/sec_bpe_blog): What happens when you run byte pair encoding (BPE) on encrypted text? An experiment inspired by Karpathy's minBPE that became a PyPI package. - [grove: I built Obsidian for the terminal](https://yash-sri.xyz/blog/grove_blog): grove is Obsidian for the terminal: a notes app that lives where you already work. Why I built it after a year of switching windows, and how it works. - [pairy: AI pair programming the way I like it](https://yash-sri.xyz/blog/pairy_blog): pairy is a Neovim plugin for AI pair programming on the Gemini API: most of what Copilot does, for free, in code you can read. ## Pages - [Home](https://yash-sri.xyz/): About Yash, highlights, and pinned projects - [All posts](https://yash-sri.xyz/blogs) - [Projects](https://yash-sri.xyz/projects) - [Contact](https://yash-sri.xyz/contact)