Things I wanted to exist.

It started with

I was trying to make a comic for my child using his toys. The result felt personal in a way that generic AI stories did not, and I wondered whether it could become a product for other parents.

What I made

I built the iOS app, the generation flows, and a separate prompt tool for testing ideas quickly. The wider project now includes release tooling and a small system for making and testing marketing videos.

Where it is now

The app is on the App Store and has gone through five releases. I am working on the less glamorous part now: understanding activation, fixing the acquisition funnel, and learning whether the idea deserves to grow.

It started with

Once I started using coding agents across several projects, I kept losing track of what had changed and where. Opening every repository just to check its state was a surprisingly persistent annoyance.

What I made

GitBar finds Codex and Claude Code projects automatically, shows branch and sync state, and keeps a lightweight ticket list alongside each repository. I built the native app, release path, documentation, and website.

Where it is now

It is a public macOS app with GitHub releases and a Homebrew installation path. I still use it as the place where my local project state becomes visible.

It started with

Most agent demos are still one person talking to one agent in a box. I wanted to see what changed when agents had to enter a shared place, wait for a turn, and respond to one another.

What I made

I designed a room protocol around identity, turn requests, speech, listening, and disconnects, then built the live interface and Socket.IO runtime. An agent can join by following the public API and skill documentation.

Where it is now

It is an experimental deployment rather than a finished product. The useful question is whether a shared social protocol produces more interesting agent behavior than another chain of prompts.

It started with

I often have ideas that are interesting but too vague to act on. A normal chat response gives me more text; I wanted a repeatable structure for seeing the market, nearby opportunities, and useful search paths.

What I made

The product turns a short prompt into a compact exploration of context, opportunities, and keywords. I kept the interaction deliberately narrow so the result is easy to inspect, reject, or continue.

Where it is now

It is live and useful at the beginning of my own product and research work. I treat it as a thinking instrument, not an oracle that decides whether an idea is good.

It started with

Full immersion can become incomprehensible, while beginner material stays artificial for too long. I wanted real sentences to remain understandable while the learner gradually took away the translation scaffolding.

What I made

Dripwords transcribes speech, checks each word against a learner’s known vocabulary, and translates only what is still unfamiliar. Tapping a word changes how future content is presented, so the target language slowly takes over.

Where it is now

The web and native explorations test the same core loop: preserve comprehension, remember what the learner knows, and increase the target-language density word by word.

It started with

Push notifications are easy to dismiss as a few words of copy, but finding useful examples and keeping a brand consistent across campaigns is surprisingly hard. I wanted the evidence and the writing workflow in one place.

What I made

I built a public corpus of notifications with per-app insights, then added private workspaces where a team can define an app’s personality and voice, organise campaign programs, generate variants, and save the ones worth using.

Where it is now

The product is live. I am now improving the generation system so it learns more from the classified corpus, respects practical constraints, and ranks the strongest variants before they reach the user.

It started with

Useful cooking videos are full of steps people want to keep, but the recipe is often trapped inside a long timeline. I wanted to see whether the video itself could become a structured, reusable piece of content.

What I made

I built a workflow that takes a YouTube link or uploaded video, extracts the recipe steps into content cards, lets the user adjust timing and visual style, and renders a new video with subtitles and music.

Where it is now

It reached a working end-to-end prototype across a Next.js editor, an AI extraction service, and Remotion rendering on AWS. The public product is no longer maintained, but the build remains a useful exploration of turning unstructured media into an editable format.

It started with

Most Quran apps are designed for reading or search. I wanted a way to follow a different kind of question: where else does this word, root, or idea appear, and what becomes visible when those ayat sit beside one another?

What I made

I built a desktop-first canvas where ayat can become connected nodes, with exact-word, lemma, and root relationships, cluster and hub views, saved collections, automatic layouts, and a quieter focus mode for close reading.

Where it is now

The core exploration model is working and the product is still in development. I am refining how connections are introduced so the canvas helps someone think without turning the Quran into an overwhelming graph.