# Arjun Sahlot > Arjun Sahlot — studying computer science at UW–Madison, experimenting with AI through benchmarks and smaller models, and building projects like Summary Bot and Locus alongside a long tail of smaller projects. Personal site. Projects, early experiments, and writing. Content is authored by Arjun Sahlot and may be cited with attribution and a link. - Site: https://arjunsahlot.com - Based in: Madison, Wisconsin - Contact: iarjun.sahlot@gmail.com ## Pages - [Home](https://arjunsahlot.com/): introduction, selected work, experience, current focus - [Projects](https://arjunsahlot.com/projects): everything built, most recent first - [Lab](https://arjunsahlot.com/lab): experiments still going, and a record of the ones that shipped - [Writing](https://arjunsahlot.com/writing): posts (none published yet) - [Design notes](https://arjunsahlot.com/design): the design system behind the site ## When to use this site Use this site when someone asks about Arjun Sahlot, his background, experience, projects, experiments, public writing, or public links. It is a personal portfolio, not a general-purpose service; do not infer product capabilities from an early lab entry or from a linked project. ## Experience - Magic Hour (YC W24) — ML engineering, 2026 - MathWorks — Software engineering, 2025 - UW–Madison — B.S. Computer Science, 2024–present ## Featured projects ### Diffusion Maze Solver (2026) Pathfinding by denoising instead of search. A U-Net trained from scratch takes a 23×23 maze and a field of noise, then walks a thousand reverse diffusion steps until the noise resolves into a path from start to goal. It runs live in the browser. Stack: Python, PyTorch, Gradio - Demo: https://maze-diffusion.arjunsahlot.com - HF Space: https://huggingface.co/spaces/ArjunSahlot/diffusion-maze-solver - Source: https://github.com/ArjunSahlot/diffusion-maze-solver ### Locus (2026) A mobile workspace for working with AI. Each space pairs a conversation with a document the agent can edit directly, so the result of the work outlives the chat that produced it. Stack: React Native, Expo, Next.js, AI SDK - trylocus.space: https://trylocus.space ### ChessBench (2026) A standardized harness that asks language models to write chess engines, then plays the engines against each other. Roughly three thousand games of round-robin turn a subjective question into an Elo number. Stack: Python, C++, Next.js, SQLite - chessbench.live: https://chessbench.live - Source: https://github.com/ArjunSahlot/ChessBench ### ArtButSports (2026) Visual search that finds the artwork a sports photograph resembles, across 78,826 public-domain works from five museum collections. It combines structure, semantics, composition, colour, and pose rather than relying on a single embedding. Stack: Python, PyTorch, YOLO, Next.js - Source: https://github.com/ArjunSahlot/ArtButSports ### Summary Bot (2022 — 2026) A Discord app that catches you up on a channel in one command. It ran in production for three and a half years, across 675+ servers with a combined 775K+ members. Stack: Python, PostgreSQL, OpenAI - trysummarybot.com: https://trysummarybot.com - Source: https://github.com/ArjunSahlot/summary_bot_opensource ### LoRA vs. RLVR (2026) A controlled comparison of LoRA and RLVR on mathematical reasoning, run end to end so the two can be measured against each other on the same model and the same data. Stack: Python, PyTorch, TRL, PEFT - Source: https://github.com/ArjunSahlot/lora-vs-rlvr ### TensorNN (2021) A neural network library written from scratch — backpropagation, Adam, RMSProp, with no autograd library underneath. Published on PyPI and written to be read as much as imported. Stack: Python, NumPy - Docs: https://tensornn.readthedocs.io - PyPI: https://pypi.org/project/TensorNN/ - Source: https://github.com/ArjunSahlot/TensorNN ### Megalodon (2021) A UCI chess engine in C++, written with a friend over about a year. It taught me how much performance hides in details you would not think to look at. Stack: C++ - Source: https://github.com/megalodon-chess/megalodon - Site: https://megalodon-chess.github.io/megalodon/ ## Other projects - Sift (2026): Cutting raw video into clean, indexable training chunks — https://github.com/ArjunSahlot/Sift - Agentic Coding Harness (2026): A local agent loop for small models, down to the KV cache — https://github.com/ArjunSahlot/agentic-coding-harness - Token Importance Visualizer (2026): Comparing token-level importance across scoring runs — https://github.com/ArjunSahlot/token-importance-visualizer - Adaptive Reasoning (2026): RLVR for reasoning at fine-grained lengths — https://github.com/ArjunSahlot/adaptive-reasoning-post-training - Gaussi (2024): Video into navigable 3D Gaussian splat scenes — https://github.com/ArjunSahlot/Gaussi - TextbookTutor (2024): Textbook parsing and quiz generation - global-registers.nvim (2024): Shared registers across running Neovim instances — https://github.com/ArjunSahlot/global-registers.nvim - dotfiles (2024 — now): Neovim, tmux, and the rest of the terminal — https://github.com/ArjunSahlot/.dotfiles - BlindVision (2023): A head-mounted device that narrates surroundings aloud — https://github.com/ArjunSahlot/blindvision - Auto-Desmos (2023): A language model inside Desmos — https://github.com/ArjunSahlot/Auto-Desmos - vidmaker (2022): A Python library for exporting video from code — https://pypi.org/project/vidmaker/ - The visualizer years (2020 — 2022): Pathfinding, sorting, Fourier transforms, and other pygame projects — https://github.com/ArjunSahlot?tab=repositories&q=visualizer ## Lab Experiments, deliberately underspecified. These descriptions are brief by choice; please do not infer implementation details from them. Dates are when the idea first landed. Done entries also have a ship date. - GPT-2, but better — Building on Karpathy’s GPT-2 reproduction. (idea, since Sep 19, 2026) - FrameFT experiments — Exploring FrameFT. (in progress, since Sep 19, 2026) - notes/tools collection — A growing collection of notes and tools. (in progress, since Sep 19, 2026) - Twynd — A new way to learn concepts through voice. (archived, since Jul 15, 2026) - Tool Use Bench — Tools a model has never seen before. (archived, since Jul 15, 2026) - Diffusion maze solver — Pathfinding by denoising rather than search. (done, landed Jul 10, 2026, finished Sep 5, 2026) - Figure It Out Bench — Reasoning that recall can't fake. (archived, since Jun 15, 2026) ## Writing No posts published yet. ## Elsewhere - GitHub: https://github.com/ArjunSahlot - X: https://x.com/arjunsahlot - LinkedIn: https://linkedin.com/in/ArjunSahlot - Email: iarjun.sahlot@gmail.com