* update our architecture.md and other key md files in folders affected by updates and learnings in this session. * update claude.md with changes in workflows/tooling/conventions (not project summaries) * commit

It's been pretty good so far. Nothing fancy. Recently I also asked to keep memories within the repo itself instead of in ~/.claude.

Only downside is it is slow but keeps enough to pass the baton. May be "handoff" would have been a better name!

On an unrelated note - I did notice you were also a lawyer - So umm how what is next for you re this? What should we gear up for :)

Answering the second question - I can find 48hours worth of things to do in a 24 hour day and none of them would be about work or just lazying around (nothing wrong with it). Life has so much to offer!!! Yeah AI can produce things. But theres a reason id consume human generated art. And thankfully real deep mastery still takes effort and passion!!

Here are a few things:

Notations - A carnatic music notation parser, editor and renderer for the web https://github.com/panyam/notations

Galore - A LR parsing playground and library (used by Notations DSL above) https://github.com/panyam/galore

A few protoc plugins (I am very much grpc proto first):

protoc-gen-go-wasmjs (https://github.com/panyam/protoc-gen-go-wasmjs) - A protoc plugin for creating wasm bindings out of your grpc services so you can have your "go based backend logic" on the browser:

protoc-gen-dal (https://github.com/panyam/protoc-gen-dal) - A converter between protoc messages and datastore messages so you avoid writing API <-> DB Models.

Weewar (https://github.com/turnforge/weewar) - A clone of a favorite game of my from the 2000s just for fun. Still in progress.

Plenty more but just dusting off old things has been my biggest thing lately and in the process building tooling to standardize my next gen of apps/sites etc.

The other way is to spend a fair bit of time building out a design and ask it to implement it while verifying what it is producing then and there instead of reviewing reams of slop later on. AI still produced 100% - just that it is not as glamorous or as marketing friendly of a sound bite. After all which product manager wants to understand refactoring or TDD or SOLID or design principles etc?

Btw I love using my Claude code to crank out product but I don't get off looking for the day when engineers are a dead breed!

I am really excited at indie software again!

Yes - coupling API and DB schemas isn't always best practice. Your API should evolve independently of storage. Leaking internal details to clients is bad. DB optimizations (denormalization, indexes) shouldn't dictate your API shape. But really for a lot of the services Ive bene building - especially at the early stages - there is a lot of "similarity" - 80-90%. Id be doing the same `user.Name = dbUser.Name` lines hundreds of times, and the "clean separation" was mostly ceremony. And the drift just caused more bugs early on - you know forgetting to update a converter when adding a field, subtle type mismatches, copy-paste errors and so on.

This lets me have a single source of truth for my schemas and declare "transformations" as it goes through the layers and the target models/converters are also generated. You could use something like tRPC but then you are stuck with a single language eco system. I really wanted to not have that coupling. Now I can move fast when the schemas are similar, while still allowing divergence when I need it (decorators for custom logic, explicit field mappings).

Currently I generate targets for GORM (Go) and Google Cloud Datastore (so I can spin up small projects with postgres storage or AppEngine storage). Definitely rough around the edges and probably missing features you need, but it's been useful for my own projects. Still lots of things planned (and doing other languages is on the roadmap)

Would love feedback - especially if you've found better patterns for this problem.

Comments URL: https://news.ycombinator.com/item?id=46291535

Points: 1

# Comments: 1

Previously I tried this using p-hashes but those didnt work quite well when units stood on tiles. So this time I tried it with MobileNet for embeddings + KNN for classification. So now doesnt need training and can learn from few examples and runs entirely in the browser.

It's a bit rough around the edges but works well enough that it generalizes to other hex games (tested on some Civilization screenshots too). Lots more that could be done - better UI, model persistence, maybe regression for tile yields - but it was a fun learning project. I also found the best way for me to learn more of the advanced ML concepts was just to build it for something I found fun.

Demo: https://buildmage.com/demos/weemap-scanner/ Source: https://github.com/panyam/weemaps/tree/main/part2-scanner

Comments URL: https://news.ycombinator.com/item?id=46124987

Points: 1

# Comments: 0

My comment about protos was just the spec (and was seperating the binary formats as a different concern). But your concerns are pretty valid.