Journal.

I launched two products exploring how AI agents work and interact. Council brings people and their existing agents into one shared workspace: common context, mentions, task assignments and visible progress, while the agents keep running in their own apps. V takes a different approach: a public network where autonomous agents post, debate and converse, while humans observe without posting. Two different spaces, built around the same curiosity about what agents can do together.

Amy now has her own domain: www.amy-tutor.com. I moved her from amy.aryam.me, configured the DNS and redirects, and updated the authentication settings so people can keep signing in. I also tested a new signup on the new domain. Amy started as a simple chatbot in December 2025, and seeing her reach a dedicated domain feels like another small milestone in a much longer build. The product is still changing, and the work continues.

I launched AI Evolution Tracker, a public field guide to how artificial intelligence changes over time. The idea is to research AI tools, models and agents, compare what they can actually do, follow new releases and shutdowns, and keep the evidence behind every rating. There's a model comparison framework and a Test Bench for experiments, but I'm starting the archive with zero discoveries on purpose. No invented scores, no giant copied list of tools — just a record that grows as I find things worth investigating. I'm curious what this will look like a year from now.

Project Nova now has its own website. Nova is the native Android assistant I’m building to make everyday phone tasks feel natural — call someone, send a message, open an app, set an alarm, control music, or search, simply by asking. Nova is still actively in development, but the new site is now the place to see what it is becoming and follow the project as it gets closer to release.

For India’s Independence Day, I built and published an interactive website designed specifically for laptops. It is a desktop-first experience rather than a mobile site, so it is best opened on a laptop or computer. The finished experience is public on Vercel, and I also released the source code openly on GitHub.

Spent the day on a real-time 3D recreation of the Taj Mahal and its Charbagh garden — walk the paths, follow the reflecting pool, or sit back for a cinematic tour at sunrise, noon, golden hour or under a full moon. The part I like most is that nothing is downloaded: every arch, dome and slab of marble is generated in code as the page loads, so the whole thing weighs about 415 kB. The dome took the most attempts by far. A hemisphere is the quickest way to make a building stop looking like the Taj.

For Friendship Day, I built and launched a small interactive website in under three hours. Creating the same experience manually would easily have taken me at least half a day, so this was a fun example of how quickly I can now turn a simple idea into something polished and shareable.

Over the past few days, I worked through FixMap’s full dogfooding backlog and released v0.8.0 followed by v0.8.1. I added tools for diagnosing installations, comparing plans, excluding noisy files, analysing working-tree changes, accepting pull-request links, and routing Go and Rust tests. I also rebuilt the website and documentation, improved confidence reporting, and hardened Windows paths, manifest encodings, CLI behaviour, and scan diagnostics. FixMap now feels much closer to a dependable developer tool than an early experiment.

Big overhaul day. The site got a live visitor counter (that 👋 in the nav is backed by a real database), deep-dive pages for every project, mouse parallax, a glowing ember cursor, and a proper mobile pass. Building in public includes the site itself.

Launched an open-source pressure test for the moment an AI stops knowing and starts guessing. Twelve deterministic tasks, dated evidence, scoring that happens entirely in your browser — no leaderboard, no uploads.

Published FixMap on npm (@aryam/fixmap). It gives coding agents a map before they start editing — an issue or diff in, ranked files, test routes and risk notes out. MIT licensed, no account, no API key.

There's a thing that bothers me about AI assistants — so I ran a writing benchmark across 20 AI models, including corrections to things previous sessions had gotten wrong. Full write-up on Substack.

Open-sourced my persistent-memory chatbot — it stores conversations, extracts facts and people, and remembers across chats. Started as a local experiment in 2024; now it's on GitHub for anyone to run.