README.md

metadata

title: BugLens
emoji: 🔍
colorFrom: gray
colorTo: yellow
sdk: gradio
sdk_version: 6.17.3
python_version: '3.12'
app_file: app.py
pinned: false
license: apache-2.0
short_description: Turn a bug screenshot into an honest, Jira-ready report.
tags:
  - build-small-hackathon
  - backyard-ai
  - tiny-titan
  - off-brand
  - minicpm-v
  - openbmb
  - modal
  - codex

BugLens

BugLens turns a bug screenshot and a one-line tester note into a Jira-ready bug report, a missing-info checklist, regression tests, and edge-case notes — using a ~1.3B vision model that stays honest about what it cannot see.

Build Small Hackathon · Track: 🏡 Backyard AI · Model: openbmb/MiniCPM-V-4.6 (~1.3B)

• 🎥 Demo video: coming soon
• 📣 Social post: coming soon

The idea

QA testers spend real time turning screenshots into structured tickets. BugLens does it in one read — and its differentiator is honesty: a small model that refuses to invent context it cannot perceive (browser, OS, device, user role, environment, backend state) and instead hands you a "confirm before filing" checklist.

How it works (technical approach)

• Frontend: a custom React UI served via gradio.Server (FastAPI + Gradio's API engine) — deliberately beyond stock Gradio (Off-Brand).
• Contract: one validated Pydantic BugReport flows through the whole app and serializes to the exact JSON the UI renders into four cards.
• Pipeline (two calls): screenshot → factual UI observation (vision) → strict-JSON structuring → Pydantic validation → cards + Jira / GitHub / CSV exports.
• Model: openbmb/MiniCPM-V-4.6 (~1.3B, SigLIP2-400M + Qwen3.5-0.8B) — image-native and strong at OCR / UI reading (Tiny Titan, OpenBMB).
• Inference backend: a Modal GPU endpoint the Space calls over HTTPS (Modal); a mock mode keeps the UI fully runnable during development.

Run it locally

See UV.md for details.

uv sync --extra dev
uv run pytest            # contract, render, and endpoint tests
uv run python app.py     # serves BugLens on http://localhost:7860

Verified stack

• Python >=3.11, package manager: uv
• Gradio 6.17.3 · Pydantic 2.13.4 · Pillow 12.2.0 · Requests 2.34.2
• Modal backend: Modal 1.4.3 + Transformers 5.10.2 (install with av instead of torchcodec to avoid CUDA conflicts)

Plans

• 01_research_verified_stack.md
• 02_step_by_step_execution_plan.md
• 03_architecture_flows.md
• 04_prompt_schema_quality_plan.md
• 05_submission_readme_demo_social.md

Contribution strategy

Codex contribution is part of the submission strategy. See docs/CONTRIBUTION_STRATEGY.md.