Surviving My First Overnight Hackathon: Local AI, Git Panics, and Zero Sleep

Surviving My First Overnight Hackathon: Local AI, Git Panics, and Zero Sleep

It is currently past 6:00 AM, the sun is rising over the NST campus here in Sonipat, and my screen is burning a permanent rectangle into my retinas. I just survived my very first overnight hackathon at Hacktoberfest Hack Day. Going into a 12-hour coding sprint from sunset to sunrise sounds romantic in theory, but the reality is a gritty mix of caffeine, broken dependencies, and sheer stubbornness to not let a complex problem beat you.

My team—Srivastava,aditya dhanraj singh, Resmanth, and I—decided to tackle a massive pain point in engineering: lost context. Developers make critical architectural decisions in messy Slack or Discord threads, and months later, no one remembers why a specific framework was chosen or rejected. Standard cloud LLMs can summarize these chats, but they hallucinate facts and leak proprietary code. To solve this from first principles, we built IssueFlowP6, a completely offline Decision Intelligence CLI that extracts decisions locally and rigorously fact-checks them.

I took ownership of the extraction architecture. My job was to write the Python logic that feeds raw developer chat logs into a local Qwen 2.5 model via Ollama, forcing the AI to output a strict JSON schema containing the final decision, rejected alternatives, and verbatim supporting quotes. Building the script and getting the raw object data out felt incredible right up until I accidentally polluted our main branch. There is a very specific type of panic that sets in at 1:30 AM when you think you just ruined your team's clean repository. Learning how to execute a hard Git reset to wipe my unstaged tracks while keeping my actual code safely isolated on my own branch was a trial by fire I won't forget.

While I was wrestling with Qwen's extraction formatting, Srivastava was building a deterministic verification engine to catch AI hallucinations, and Gagandeep was wrapping our scripts into a clean command-line interface. We integrated everything just as the overnight exhaustion was hitting its peak. Running the final terminal command and watching the pipeline successfully parse the text, intercept two hallucinated quotes, and spit out a perfectly verified JSON file made every single hour of sleep deprivation completely worth it.

I am writing this right before final evaluations to participate in the Hacktoberfest blog challenge, where sharing our project experience and posting it on DEV gives us a shot at a Rs 1,000 prize. The code is locked, the pitch deck is styled, and we are about to present a fully functional, privacy-first local AI tool. Regardless of the final judging, building this end-to-end pipeline from scratch in a single night is a massive win. Now, it is time to submit the blog link and finally get some sleep.