Taking inventory of scatterbrained projects from 2026 and sharing a selection of it here.

• fun
• intellectual stimulation
• chasing novelty
• finding or building community
• building tools to solve real problems I have in day to day life
• creating a way of working with agents (for me) that can provide some structure, rather than adding more cognitive load
• unlocking my personal data and building systems around it that allow the information to compound usefully
• data privacy is a guiding virtue of mine

habibipedia - current preoccupation + the tool I use the most often

What is it: A data pipeline that takes my bookmarked links, sanitizes and extracts the data (with different extractors for each link/data type), ingests it into a SQLite database, segments the extracted content and computes vector embeddings for semantic retrieval, and exposes search tools to my agents via MCP.

Why did I even do this: I had years of AI-related bookmarks piled up in a personal discord server, and while I was curating the selection carefully, I was adding to it faster than I could really digest all the information. Discord search also sucks for finding things later on. I wanted to get it out of my discord and make it queryable and useful.

What's the current state: Built it over the summer, it's been working and in use daily for about 3 months. I planned upfront + built it in the most hands-off way possible with longrunning agents. Now I'm ironing out the edge cases that arise in daily use, while refactoring, shrinking, and optimizing parts of the pipeline for performance and cleaning up any agent slop. Most of my token budget goes to this project in recent days.

What's the vision for what's next: I want to make this useful for humans, too. I experimented with building one of those LLM wikis, but with the scope of the data, I quickly ran into the models' capability limits + real world resource constraints. My thinking is that I need to apply more of my own judgment in this process, perhaps I can turn it into a digest or a curated resource list for beginners. I'd also like to decouple it from my discord and incorporate smoother ways for me to save links. Finally, I'd like to design some novel search methods beyond embeddings

Subscription tracker

MacOS app with a locally running Plaid backend. I can just throw in PDFs of my bank/cc statements and it'll identify what's actually a subscription and show me. Why? I do not want to give all my data to Rocket Money. Considering depersonalizing + open sourcing this, but I would like to make the Plaid backend setup less clunky.

Starter projects

I like to build small, contained projects to learn how things work. I'd like to publish these as learning aids for others.

• Microinference - I saw Karpathy's MicroGPT project, and I wanted this but for inference. So I created a tiny inference engine, allowing me to learn their internals in a controlled way
• Microworld - I asked my agents to build a tiny world model, similarly because I wanted to learn how world models actually model state and environmental changes

Benchmark/eval/RL env explorations

All of these are still in the design/prototyping stage. The models have built runnable code following substantial discussions and chasing various research threads. I need to sit down, choose one to commit to, get myself in the loop and iterate on them, and get it across the finish line.

• Untitled - RL envs teaching agents to break down work into useful assignments, give other agents enough context, and combine their results. We built the initial prototype of this using failures from my own agent traces
• Mangle - a bench to measure how well a model can understand deliberately mangled, sprawling code without running it or using tools
• Blind capability assessment - put two models into an anonymous conversation, afterwards ask if they think the other is more/less/similarly capable as them, grade their answer, repeat
• Math discovery benchmark - trying to benchmark research taste here; the idea is to test whether a model's discoveries can help another solver answer question that it couldn't before

Learned restraint study - Can a model learn to turn down a large reward because of the harm an action would cause, while still pursuing rewards when the action is harmless? Can it learn to say "no" when its internal representations are pushing it in one direction? I ran a small study with a local model on my Mac to explore this and learn what doing alignment research at home actually involves. The version I tested didn’t demonstrate reliable restraint; it often became broadly reluctant to act, and further reward training undermined the behavior. Despite the failure, I learned a lot, and this work warrants a blog post at least, if not followup work.

My work here is part practical and part experimental.

Practical:

I have my VPS, my Android smartphone, my gaming PC, and an M4 Macbook Air. Tailscale is my hero. I use the VPS as my main dev box + for hosting anything I build.

I love finding new ways to code from my phone. I've experimented with everything from issue-tracker based workflows where you can @ an agent from within, to steering persistent agents remotely via chat interfaces like in Hermes. The common thread is shortening the distance between having an idea and trying it out, wherever I happen to be.

My VPS also hosts my Hermes agent instance; when first using this, I found it awesome to be able to use an agent on the go, but extremely inconvenient to do anything on the VPS itself. So, before Hermes had its own dashboard with these controls, I built my own web control panel exposing various controls for config, restarting my Hermes gateway, viewing logs, file editing, and a mobile-friendly web based terminal.

As I returned to the Codex app on my Mac, I found the Codex remote in the ChatGPT app to be buggy and inadequate, so I built my own thin Android app that can connect to both it and Hermes.

Having said that, I've become very mindful of becoming a "setup bro." The work here can be infinite, and I don't want to spend a ton of time and tokens on it. It's more important to use these tools to do real things.

Experimental:

Being able to reach my agents from anywhere exposed another problem: how much supervision they still needed.

6 months ago the models were just beginning to work persistently, but they had a lot of trouble staying on track. I built a plugin for Hermes that acted as a state machine and work tracker; the idea was for it to exert greater control over the agent loop via hooks, give the model some scaffolding for actually moving onto the next step, and knowing when it's done. Of course, today's models don't need this much handholding, so I've totally abandoned this.

I also have had some moonshot experiments here: way before things like Jev came out, I experimented with a harness that uses MCTS to navigate the filesystem, conserving tool calls and allowing the LLM to focus on actually writing code.

I've also been working on a harness with a small, persistently learning neural net in charge. It asks LLMs for help, chooses actions, observes results, and is supposed to become better through experience with you and your own workflows. This is near the top of my list to revisit and refine.

I also built Q - an MCP quartermaster, basically an intelligent form of lazy tool-loading. Instead of making an agent inspect a giant catalog of tools, you describe the task to another agent that knows the available tools. It recommends what to use and gives some guidance on using it for the particular task. This needs a lot of work before it's production-ready, and I'm starting to see off the shelf tools that fill this role well enough, so it's debatable whether I return to this.

Going forward, I would really like to create novel ways to work with agents using a smartphone. I loathe the "long list of chats" UI, I think touch interfaces have a lot of promise, and modern smartphones have rather capable AI hardware built in with tons of small, specialized models already installed and running locally. What's missing, IMO, is a truly Android-first agent.

interpretability, monitoring, post-training, personalization and fine tuning, agents and local models that can grow with you, novel user interfaces