woke up super groggy. i've been sleep deprived for the second day. got to baincapital at 9 a.m.. the intial phase felt like we were going in circles, the idea still wasn't clear, things weren't working, it feels like you're navigating in the fog.
so many issues, instructor with claude was too slow, concurrent threads were the way. to speed things up, we used llama3 on pure groq api, but litellm had better abstraction. then we used llama hosted on cerebras because it was almost 10x faster. then there were rate limits and max tokens per day. there was also issues with generating json, so num_retries=1000 was the fix. then the responses weren't great, so it was gpt4o because claude 3.5 sonnet doesn't support json mode on lite llm.
this hackathon was a much better experience. no staying up the entire night to code a streamlit app alone. i knew i could count on my teammates, my last one felt like i was the only one rowing the sinking boat. we also had cursor to help out, i barely wrote any code by hand. in hackathons, you optimize for speed and execution, the fastest to debug and ship out the idea wins. i suppose if you're efficient with AI tools, and your goal is to just build, it's understandable. but i feel like i lack strong programming skills, it made me want to join recurse center. maybe sometime in the future.
nearing the deadline, watching the fog clear, the slides coming to shape and the demo working is one of the best feelings ever. vercel ceo came over to our table and we were running on over 400 doc files of vercel, it was so fun nerve-wracking but exciting to watch it run, especially with minutes left. everything aligned so well.
some ideas from the hackathon:
- pca cluster next.js github issue and chat
- chrome extension to record screen and audio to generate browserbase functions for agents
- ai powered feature flag management
- generate API docs in minutes
- AR validator with openCV, mediapipe, and openai
- ai categorization of changelog
- education tool for learning how github repo works
- spelling and grammar checker for docs (our team)
- minusX: chat assistant for jupyter notebooks, tableau, etc.
- hands off anomaly detector, replace datadog
- error logs fixer
- merge databases with LLMs
- tinder x arxiv
- blog post generator from new features
- codebase -> graph -> chat
- intent classification of chat
- better reasoning and planning for llms
- data extraction with llms
we won the best demo prize (all credits to A for his awesome demo), but we had to rank the pool of prizes, and in the end we only got one oura ring for 4 of us. the prize was disappointing, but nothing beats the experience. even though there are 3 quizzes, practicum rankings and interviews coming soon, hackathon experiences like this is why i wanted to come to sf.
i love the hackathon energy. i was creating a projects ideas page, as i was cleaning up the open tabs in my arc browser. i want to keep this energy alive everyday, building fun things that i would use, and share them with the world.