Writing
Things I have had to learn twice, written down so I only have to learn them once more.
22 posts
24 Sept 2026 · 8 min
TypeSafe's new Jev model is pitched as a fast alternative to LLMs for 'smart if-statements' - classify, route, score, branch. That framing is closer to the NLTK code I wrote in 2017 than to what a chat model does, and tracing the line between them explains why most software problems were never conversations to begin with.
ReadIn 2019 I built a chatbot for my college's placement cell on Dialogflow, before anything like ChatGPT existed. Revisiting the code, the thing that stands out is how cleanly it split the job: a trained model recognized what you wanted, and code I wrote by hand decided what to do about it.
22 Sept 2026 · 9 min
The last part of the Jellyfin series. Turning the bot script into a systemd service, keeping the local Telegram API server alive with Docker's own restart policy, log rotation, and a short security checklist for the whole setup.
21 Sept 2026 · 4 min
Part four of the Jellyfin series. The actual bot handler: turning the path a local Telegram server hands back into a file in Jellyfin's library, a permissions bug that only shows up on the first real file, and getting a Jellyfin API key to trigger the rescan.
20 Sept 2026 · 6 min
Part three of the Jellyfin series. Creating a Telegram bot with BotFather, hitting the 20 MB download wall almost immediately, and running your own local Telegram Bot API server so files up to 2 GB are just a path on disk instead of a capped download.
18 Sept 2026 · 6 min
Part two of the Jellyfin series. Pointing a domain at your Lightsail static IP, putting Caddy in front of Jellyfin for automatic HTTPS, and closing the port you no longer need open to the world.
16 Sept 2026 · 4 min
The first of a five-part guide to running your own Jellyfin server and wiring a Telegram bot into it. This part covers what Jellyfin is, setting up an Ubuntu instance on Amazon Lightsail, installing Jellyfin, opening the right ports, and getting through the first-run wizard.
14 Sept 2026 · 7 min
After a year of building software with these models I could not shake the feeling that everything was a rearrangement of something. So can pattern matching find anything genuinely new, and is that what we do anyway? Working through Cantor, the halting problem and Gödel to find out.
16 Aug 2026 · 16 min
The final part of the beginner's guide. 'Train Ollama on my own data' almost always means one of three cheaper things. This part walks through all three, in order: a one-line system prompt in a Modelfile, retrieval so the model answers from your documents, and actual fine-tuning with the tools that do it and the path back into Ollama.
07 Jul 2026 · 5 min
Part four of the beginner's guide. Ollama has been a web server the whole time. This part covers calling its HTTP API from curl, Python and other tools, keeping a model loaded so the first request is not slow, running several models at once, exposing it on your network safely, and setting the environment variables that control all of it on macOS, Windows and Linux.
19 Jun 2026 · 5 min
Part three of the beginner's guide. The Ollama library is a curated slice of what exists. This part covers pulling models straight from Hugging Face, importing a GGUF file with a Modelfile, what makes a model compatible with Ollama in the first place, and how to read the error when one will not load.
04 Jun 2026 · 5 min
Part two of the beginner's guide. How to read a model name like llama3.1:8b-instruct-q4_K_M, what a pull actually downloads, why an 8-billion-parameter model is a 4.7 GB file and not a 16 GB one, and how to check whether a model will run on your machine before you download it.
21 May 2026 · 6 min
The first of a five-part beginner's guide to running AI models on your own computer with Ollama. This part covers what Ollama actually is, how to install it on macOS, Windows and Linux, running your first model, and where everything lives on disk.
06 May 2026 · 7 min
What actually happens when you pull a model, why the file is smaller than the parameter count suggests, how to work out whether your laptop can hold it, and the settings that matter once it runs.
11 Feb 2026 · 8 min
A query and transformation language for JSON that lets you store integration logic in a database row instead of a deployment, and reshape data from many systems into one output.
09 Dec 2025 · 8 min
The DOM is a tree of nodes, so a whole browsing session serializes to plain text. rrweb does that well enough to build on: I used it to stream physics simulations to classrooms, it is what PostHog session replay runs on, and it turns a bug report into a string you can re-render pixel for pixel anywhere.
07 Oct 2025 · 7 min
For years the best programmers were the best searchers. That skill has moved, and most of what made someone good at it still applies.
19 Aug 2025 · 5 min
How an inverted index actually works, why relevance is a tuning problem and not a feature, and an honest comparison of the three engines most teams end up choosing between.
04 Mar 2025 · 8 min
My first blockchain project. Why a permissioned chain instead of Ethereum, what chaincode actually is, why the personal data never touched the ledger, and the design decisions I would defend today.
22 Sept 2020 · 15 min
Hedge won a Facebook productization grant in 2017. It captured browsing activity in internet cafes, scored it with NLP, and moved it over a handshake I built myself. Most of what I learned came from the parts I got wrong.
14 Jun 2020 · 9 min