I understand the basics of programming, but these days I don't write code myself. What I do: think through what a system should do, which parts it needs, how they talk to each other and where the boundaries are. The implementation is done by AI, and I check whether the result holds up.
I'm saying this openly because it's how these projects actually came about, and because honesty matters more to me than a polished image.
An honest disclaimer first
These are learning projects. I use them to find out how far you can get with AI when you take care of the thinking and the structure yourself. None of them is in real use, none has been reviewed by an independent developer, and I can't vouch line by line for code I didn't write myself. For each project you'll find what it is, what it can do and, just as important, what it isn't.
Learning project · not in use
Log Gateway
A bouncer for log data, written in Rust
What it is
Programs send their log data to the gateway. Before anything is stored, it removes personal data such as email addresses, IP addresses or IBANs. After that it counts and limits requests per customer and stores the data compressed. It started as part of a larger experiment in monitoring internet routing (BGP). Later I split it out so it could stand on its own.
My part
I decided which building blocks exist, how customers (tenants) are kept apart and which metrics matter. I also decided to split the gateway out of the larger project instead of dragging it along.
What it can do
Detect and redact personal data using 7 fixed patterns
Limit, count and apply quotas to requests per customer
Store logs compressed and export them to S3, plus metrics for Prometheus
Around 140 automated tests, written by the AI alongside the code
What it isn't
Not in real use: there are no users and no running server.
No reliable performance numbers. Earlier benchmarks came from a server that no longer exists and don't apply to this version.
Not a complete privacy solution: 7 patterns miss plenty of personal data.
Not reviewed by anyone else: no external code review, no security audit, no licence yet.
A Bloom filter answers "have I seen this before?" very quickly and with very little memory. It can sometimes be wrong, but only in one direction. BloomScan splits this into nine small modules. The interesting part here is less the library than the way it was built.
My part
I designed a small organisation of AI roles: a coordinator that writes no code itself and only hands out tasks and runs tests, four specialist architects (interfaces, logic, data flow, integration), a chief architect for conflicts, and one agent per module that may only change its own file. Every role has fixed read and write permissions and communicates only through defined files.
What it can do
Show that AI work can be structured with clear roles and permissions
Store and query entries, monitor how full the filter is, compare two filters
Split text into word groups (n-grams) and add them to the filter
Around 90 automated tests across 10 modules
What it isn't
Not a finished library: unpublished, no documentation for outsiders, not used anywhere.
Nothing new: Bloom filters have existed since 1970, and mature libraries already exist.
Not proof that the agent workflow works in general. It's a single experiment on a manageable problem.
From idea to finished episode: script, voices, mastering, video
What it is
A pipeline for whole podcast series: Claude writes the scripts, a locally running speech model (Qwen3-TTS) voices them, and each episode ends up as a finished MP3. Lolfi puts the episode under a calm lofi video.
My part
I designed the formats and the flow, with five templates from classic narration to crime audio dramas to a language course. In crime mode every character gets their own knowledge: what they know, what they hide and what they wrongly believe. The aim is for contradictions to come out of the story itself instead of just being claimed.
What it can do
Turn a series idea into episode plans and scripts, reviewed by a second AI pass
Voice multi-speaker audio dramas and even out loudness (mastering)
Output sound-effect cues as a timestamped list for editing in an audio program
With Lolfi, stretch a short video loop to episode length and combine it with the audio
What it isn't
Not a product for others: it needs a Claude subscription, a configured machine and technical know-how to set up.
Not fully automatic: I create images and video loops for Lolfi by hand in other tools, and sound effects are added afterwards.
No guarantee of good stories: quality depends on the AI and varies. Listening and picking is still manual work.