This is a submission for the GitHub Finish-Up-A-Thon Challenge
What I Built
CodeAutopsy. You know that feeling when you clone a repo and just...
stare at 47 folders with no idea where anything is? That's what I
built this to fix.
Paste any GitHub URL. It parses the entire codebase, builds a
dependency graph, and tells you exactly what imports what, where
execution starts, and what breaks if you touch a specific file.
Started this a few months ago because I was tired of spending the
first two days of every new project just figuring out where
everything lives.
Demo
Live: codeautopsy-lyart.vercel.app
GitHub: github.com/Sidhant0707/codeautopsy
The Comeback Story
Shipped the core, posted on Reddit, someone in China tweeted it in
Chinese and it got 7,400 views overnight. 66 stars, 262 unique
cloners in two weeks. All organic. I was pretty stoked.
Then everything broke.
The AI layer was sending 40-60k tokens per analysis to a single
provider. Groq hit its 100k daily limit in 18 runs. Switched to
Gemini — project got permanently banned. Tried Cerebras — they
dropped Llama entirely overnight. SambaNova had a 20 requests per
day limit which lasted about 4 minutes.
Two full days of provider roulette.
The fix was rethinking the whole approach — splitting the single
massive AI call into focused smaller calls across multiple providers,
building a fallback chain so one 429 doesn't take the whole thing
down, and stabilizing the streaming JSON repair so users never
stare at a blank screen mid-analysis.
The interview mode feature I'd built — where an AI interviews you
about the codebase while nodes on the graph light up — got parked
on a branch until the core was stable enough to actually support it.
My Experience with GitHub Copilot
Copilot was open the whole time in VSCode. The place it actually
helped most was the partial JSON repair algorithm — I'm streaming
incomplete JSON token by token and trying to parse it in real time.
The edge cases are nasty: unclosed strings mid-token, mismatched
brackets, empty chunks. Copilot kept catching the ones I missed.
Also saved me a lot of time tuning the d3-force physics for the
dependency graph. Getting node collision and link distance to feel
right took a lot of iteration and having inline suggestions for the
parameter values was genuinely useful instead of just guessing.
Top comments (0)