← Back

The interest in automation came long before generative AI. I had always looked for the repetitive part of a process and tried to take it out. What generative AI changed was the speed. It gave that instinct the gas it had been missing, and things that would once have taken a month of evenings now take a weekend.

On client work, my team and I use agentic coding tools with an MCP layer wired into the systems the work already lives in: the ticket tracker, the knowledge base, the shared repositories. The point is not code generation. It is that the agent can see the same context we can. It changes the early hours of a problem most. Instead of reading through undocumented customisations by hand, the agent maps the codebase and explains how the pieces connect. Instead of moving between the code, the logs and the ticket separately, it holds all three at once and points at the fault.

Why I build my own

Most of what I need already exists as a product, and most of those products do far more than I want. I usually need something minimal: one job, done the way I actually work, with nothing else in the way. That version is rarely in the market, so I build it myself.

Things I Have Built

None of these are publicly available yet. They were built for my own use and carry my own assumptions. Generalising them into something another person can pick up is its own piece of work, and one I plan to take on.

Hebrew Learning Application

React · Node · Claude API · self-hosted

Spaced repetition and active recall over a vocabulary that builds out of real texts rather than a fixed word list. A pipeline splits and aligns sentences, adds vocalisation, and transliterates; clicking a word pulls a lexicon entry, and verse search runs against an open scripture API.

Nutrition Tracker

React · Node · SQLite · Claude API

Free text in, structured nutrition out. You write what you ate the way you would say it and a model resolves it into quantities and macros. Energy balance is modelled properly, with resting expenditure separated from logged exercise rather than collapsed into a single number, and daily targets are versioned so that changing a goal never rewrites history.

Calendar and Reminder Service

Node · CalDAV · Google Calendar API · Twilio

Consolidates three calendars, two work and one personal, and sends a single morning summary over WhatsApp, with reminders scheduled ahead of each event. Runs unattended on a fixed timezone schedule.

Transcription and Note Agent

Node · React · AssemblyAI · scheduled agent

I would rather read than listen, so this collects podcasts and sermons, transcribes them, and turns them into structured notes in a notebook app I built for the purpose. It is responsive, so I can pick the notes up from any device.

On the pipeline

A research workflow environment

In development

Four years of doing a PhD showed me where the workflow actually leaks: literature tracking, note organisation, keeping a citation honest to what the paper really claims, and the handoffs between all of them. The tools for each piece already exist and most of them are good. What does not exist is the connective tissue. So this is not an attempt to rebuild any of them from scratch, but to wire them into one workflow a research student can live in, from the first search through to submission.

A minimal second brain

In development

I read more than I retain. Papers, podcasts, half-formed ideas at eleven at night: most of it goes in somewhere and very little comes back. The usual fixes did not hold. Obsidian wants you to maintain the system before it does anything for you, and Notion is fine right up until there is real volume in it. What I want is smaller than both. Things go in without ceremony, get sorted by category, and come back later carrying connections I would not have drawn myself, surfaced on a spaced-repetition schedule and put to me as questions rather than shown as notes. A notes app you never reopen is only a tidier way of forgetting.