What I'm Working On

I’m a husband and father of two young kids. I spent the past few years co-founding a startup, and I now have a day job building voice AI agents at a large enterprise software company. Either way it’s a full plate, so my open-source side projects get whatever scraps of attention I can carve out during evenings and weekends – but the backlog isn’t getting shorter.

I maintain several projects and there’s always something I “owe” someone: a bug fix, a feature, an enhancement. This page exists for two reasons:

1. For you – so you can see where your bug fix / feature request sits relative to everything else, and understand why estimates slip.
2. For me – an honest prioritized list so I can stop staring at open tabs and actually make a dent.

The ordering within and across projects reflects how I currently think about the work. It shifts. I’ll try to keep this updated when it does.

Priority Queue

Project sections are in rough priority order – liwords is most active right now, then Macondo, then the others. That doesn’t mean I won’t touch Macondo until liwords is done; I switch between them. The Aerolith and ESB sections get little love at the moment.

Estimates are left blank intentionally – when I fill them in they’re usually wrong anyway. The note on each item explains context or blockers.

By Project

liwords / Woogles

Woogles is the online crossword board game platform I co-built. Real-time games, annotation and analysis tools, tournament infrastructure. The codebase is woogles-io/liwords – a Go backend with a TypeScript/React frontend.

This is the most active project right now. The main threads:

Puzzles. The puzzle system exists but is pretty bare – no tags, no filtering, no structured search. Overhauling this depends on code I very recently merged into Macondo, so it’s unblocked and near the top of the queue. (#1706)

New languages. Slovene (#1845) is done. New languages also require changes to word-golib, the shared lexicon library I maintain that both Macondo and liwords depend on — so each one lands in two repos.

Infrastructure upgrade. Postgres, NATS, and Redis all need updating to current versions. The catch: any serious upgrade means downtime, and downtime means I need to ship adjourn-game functionality first so live games can survive a restart. So the infra upgrade is gated on that feature.

Game architecture overhaul. This is a large refactor being organized using the Mikado method – a technique for untangling complex dependency graphs where you try a change, note what breaks, back it out, fix the prerequisite, and repeat until the path is clear. The goals: remove the in-process game cache (which currently forces hard-cutover deploys and blocks running more than one API node), port the referee/game-rules logic into a new pkg/game/ package inside liwords, and drop Macondo as a liwords runtime dependency. Finished games would also move to S3, keeping the database lean. Some of this is already in flight.

Broadcast features. Player records in the OBS URL options (#1855) is shipped. Two issues remain for the broadcasting/OBS integration: native tournament mode (#1856) and more slot types (#1857).

Analysis. Two related items: allowing players to analyze a game in the browser turn by turn (using the existing server-side analysis infra, no WASM needed for now), and moving stored analyses off the database and onto S3 to keep the DB lean.

And then there’s just a long tail of smaller issues in the repo that each feel tractable in isolation but add up.

---

Macondo

Macondo is the crossword board game engine – move generator, equity calculator, endgame solver, and bot. It powers analysis and bots on Woogles and is probably the strongest open-source engine for this class of game.

New lexicons. Adding Swedish (#491) is the concrete next item. It requires changes to word-golib (the shared library for lexicon data and letter distributions), the same way Slovene did. The word-golib bug where lexicons with unrecognized prefixes threw errors instead of being handled gracefully (#496) is fixed.

Bayesian inference. I’ve been doing research into improving the inference module – searching mathematically for the best value of τ (the softmax temperature that controls how rational we assume the opponent to be when computing P(play | leave): lower values assume near-optimal play, higher values allow more weight for sub-optimal moves) and running related experiments. This one is slow-burn research rather than a concrete deadline.

Next-gen static evaluator. I trained a CNN-based static evaluator for Scrabble that already beats HastyBot’s long-standing equity engine – details in this post. But that model was trained on HastyBot self-play data, which means it learned from a bot that is itself imperfect. The next step is to collect a much larger corpus of higher-quality self-play data, explore better training algorithms (Q-learning is one candidate), and retrain to see how much further we can push the eval. The long-term horizon for this is more ambitious: convert the trained network into a NNUE-style architecture so it runs fast enough to replace the static evaluator inside Monte Carlo search entirely. If that works it would be a meaningful leap for Scrabble engine strength – the same jump chess engines made when Stockfish adopted NNUE.

GUI. Eventually Macondo should have a proper graphical interface so it’s usable without a command line (#35). This is in the “idea” bucket for now – I want to do it, I just haven’t figured out the right approach yet.

Repo: domino14/macondo

---

Aerolith / WordVault

Aerolith is my oldest project, started in 2007 as a Qt desktop app for studying Scrabble word lists. It’s been in maintenance mode for a long time. When a new official lexicon drops I’ll spend a few hours to get everything current, but it doesn’t get sustained attention.

WordVault is a study-card system I built on top of Aerolith using the FSRS spaced-repetition algorithm. It shipped, people use it, and then I didn’t touch it for about two years. It’s slowing down – queries are getting sluggish, and I need to revisit how much guess history we’re retaining. This isn’t a feature; it’s a reliability issue that needs a pass before it bites harder.

Infinite Wall mode is out, and it has accumulated a list of bugs that need a cleanup pass. Beyond that, I have ideas for other new Wordwalls modes – a Tetris-inspired “Tetrolith” variant among them. Those are still in the idea stage.

Repos: domino14/webolith · domino14/word_db_server · domino14/aerolith-infra

---

ESB – Electronic Scrabble Board

A hardware project I’m committed to delivering within the next year. It’s still in the ideation / pre-prototype phase – I don’t have a repo to link yet. The concept: a physical electronic board that can be used to play and track Scrabble games, bridging the gap between over-the-board and online play. More details to come as it takes shape.

---

Other / private

There’s a collaborative project where I’m licensing a game from a partner who also handles video creation and advertising. I will share more details in the near future. It competes for the same hours as the open-source work, which is part of why “a few weeks” can stretch into months.

---

Last updated: August 23, 2026. This page’s source is in github.com/domino14/cesars-blog.