naps62.eth

Rustacean | Building @ethuidev
@subvisual | @QuillFi | @spearbitDAO

Mastodon: [email protected]

[naps62.eth @naps62]

@CryptoLisboa Choosing the model / embeddings pipeline, tuning the prompts, and manually reviewing a few runs until I was happy with the summaries

Still not perfect though, will be ongoing work

[naps62.eth @naps62]

I built an app that turns "Diário da República" into plain language🇵🇹

See what changed, for whom, and when, without sifting through pages of legalese.

Summaries run on Amalia, Portugal's new open LLM.

https://t.co/HcduSp4Mnc

[naps62.eth @naps62]

@spatial_jonaz The gov page uses an API (it's an Outsystems page), so no scraping was needed so far
Not real time, just checking a few times a day for new articles.
they're published once a day, and take only seconds to summarize, so no need for anything fancier

[naps62.eth @naps62]

@JC48R4L Elixir / Phoenix; postgres; llama.cpp. tudo no meu home server
Cloudflare / Opnsense / crowdsec / traefik para infra e segurança.
E ligação a um openai-compatible provider para modelos externos (neste momento o https://t.co/zw01lvR55e)

[naps62.eth @naps62]

@bitsaitama @PCMagalhaes Possivelmente esse site da assembleia e o mesmo que eu encontrei quando tive a mesma ideia e vi que já existia, por volta do mesmo tempo que comecei esta 😅

[naps62.eth @naps62]

@bitsaitama @PCMagalhaes Tenho um home server com uma RTX 5070 Ti 16gb. Os embeddings + amalia correm lá. Cada artigo fica processado em ~10 segundos

[naps62.eth @naps62]

@ultravioletpet Em horas humanas, talvez ~30h, quase tudo via claude

A maior parte do tempo foi a criar e iterar specs/issues, review de codigo critico, e testes

Estatisticas reais: "107 sessions logged, 17 active days, 822 human-typed messages"

Muitas destas sessoes eram 99% autonomas

[naps62.eth @naps62]

@danielarpm Isso e precisamente a idea sim
Mas concretamente deixar os users falar com AI nao se resolve so com premium. preciso na mesma de ter hardware capacity para corresponder a escala

[naps62.eth @naps62]

@danielarpm noted. boas ideias ai, obrigado

sobre expor algum tipo de AI ao user, ja se entra num problema diferente a nivel de escala e custos.

Pelo menos enquanto isto depender da minha propria infra (leia-se: o meu GPU individual a tratar de tudo sozinho de momento)

[naps62.eth @naps62]

@IamRedListed @cburgdorf I should also stress: information loss is there *by design*. I'm compressing multiple pages of text into a couple sentences

The goal here is not to provide you full info, but just a hint for "should I go read up on this specific law? is it relevant for me?"

[naps62.eth @naps62]

@IamRedListed @cburgdorf I would guess any information loss from pt->en is dwarfed by the legalese-pt->common-pt translation

and again, anyone who needs english will already have to use a translation tool. the same problem exists there

[naps62.eth @naps62]

@IamRedListed @cburgdorf I would imagine a translation layer is probably not that hard (famous last words, I know). it was a solved problem way before LLMs

But I don't think it's a top priority right now, especially because browsers already do that for you very well

[naps62.eth @naps62]

@oleg_fem Seems I had an issue with the "http -> https" redirect, and safari was smart enough to still redirect. a toggle missing on cloudflare

should be fixed now!

[naps62.eth @naps62]

@GustavoRSSilva Para já tens RSS feed disponivel (global, e por tópico)

Eventualmente queria subscrições mais personalizadas. e.g. receber email sempre que algo novo fizer match a uma pesquisa definida por cada user

Dependendo do # de users, a escala ai pode ser um problema

[naps62.eth @naps62]

Oh, and of course, everything is OSS and self-hostable, as usual:

https://t.co/tfEoFwrFTv

[naps62.eth @naps62]

Today we're releasing an internal an internal tool we've been using for the past months

https://t.co/2E73i8GEH4

[naps62.eth @naps62]

While the initial current premise may be simple, we feel this could eliminate a lot of boilerplate and time spent on "just getting up and running"

Plus, this is only the beginning. we have much more planned to improve Deveoper Exp even further