angarrido

5 ms·

1. 1.▲
   by angarrido 5mo ago

   must people think it’s just GPU cost. In practice it’s coordination: model latency variance + queueing + retries under load. You don’t scale linearly, you get cascading slowdowns.
2. 2.▲
   by angarrido 5mo ago

   I built this because I couldn’t use Copilot or Claude at work without risking exposing internal code. This runs fully local (Gemma 4 26B), indexes your codebase, and answers questions about it without anything leaving your machine. Still ea
3. 3.▲
   How HN: AndesCode – local AI coding assistant that runs offline

   1 points by angarrido 5mo ago|2 comments
4. 4.▲
   by angarrido 5mo ago

   Local inference is getting solved pretty quickly. What still seems unsolved is how to safely use it on real private systems (large codebases, internal tools, etc) where you can’t risk leaking context even accidentally. In our experience tha
5. 5.▲
   by angarrido 5mo ago

   Awesome idea, usually LLMs lack of creativity, so layouts look mostly the same, I could give a try
6. 6.▲
   by angarrido 6mo ago

   Interesting approach. One thing I’ve been running into is that even with good alerts, the harder problem ends up being when to actually be exposed. You can get the signal right and still underperform just because of timing and volatility.