## Inspiration
Every developer knows the feeling — production is down, the terminal
is screaming a stack trace, and you spend **hours** asking one question:
> **"Where did this break, and why?"**
No tool connects runtime errors to the exact commit that caused them — intelligently and instantly. That gap inspired **Error Archaeologist**.
---
## What It Does
Paste any runtime stack trace → get the **exact GitHub commit** that introduced the bug, the **code diff**, a **root cause explanation** from GPT-5.5, and an **auto-generated regression test** — all in under 500ms via **Fireworks AI**.
---
## How We Built It
| Layer | Tech |
|-------|------|
| AI Analysis | GPT-5.5 + OpenAI Codex |
| Fast Inference | Fireworks AI |
| Version Control | GitHub REST API |
| Frontend | React + Vite + Tailwind |
| Backend | Node.js + Express + SQLite |
---
## Challenges We Ran Into
- **GitHub rate limits** → solved with SQLite caching layer
- **Diverse stack trace formats** → Codex handles all languages semantically
- **Slow inference (8–12s)** → Fireworks AI cut it to **under 500ms**
- **GPT hallucinations** → added confidence scoring; humans verify the diff
---
## Accomplishments We're Proud Of
- ✅ **90% accuracy** on 20 real-world repos in testing
- ✅ Full pipeline working end-to-end, not a demo
- ✅ Language-agnostic error parsing via Codex
- ✅ Auto-generated regression tests that prevent recurrence
---
## What We Learned
- Codex understands stack traces far better than any regex parser
- Prompt engineering took **11 iterations** to get reliable structured output
- Speed isn't optimization — with developer tools, **speed is the feature**
---
## What's Next for Error Archaeologist
- 🔌 **VS Code extension** — analyze errors without leaving your editor
- 🤖 **GitHub Actions** — auto-analyze every failed CI test
- 💬 **Slack bot** — post root cause directly to your team channel
- 🔮 **Predictive mode** — flag risky PRs *before* they merge
Log in or sign up for Devpost to join the conversation.