# 🏆 Build Small Hackathon — Submission

**Project:** UX Crime Scene 🔎
**Author:** [@kasbsquall](https://huggingface.co/kasbsquall) (Kevin Alexander Soto Burgos) — solo

---

## Track
🍄 **An Adventure in Thousand Token Wood** — build something delightful that wouldn't
exist without AI. The model **is** the detective: remove it and there's no case.

## Deliverables
| | |
| --- | --- |
| 🔎 **Space** | https://huggingface.co/spaces/build-small-hackathon/ux-crime-scene |
| ▶️ **Demo video (trailer)** | https://youtu.be/6u58YIEPrkA |
| 📹 **Full walkthrough** | https://youtu.be/WyQbY0XJ_9E |
| 📱 **Social post** | https://x.com/p36649/status/2066277845567930447 |

## One-liner
A film-noir detective investigates your UI as a crime scene — drop a screenshot and
**THE INSPECTOR** circles every UX flaw and files a verdict with a letter grade.

## The small-model build
- **Model:** `Qwen2.5-VL-7B-Instruct` — 8.3B, comfortably under the 32B cap, the
  strongest open VL model for bbox grounding at this size
- **Runs on:** **Modal** (vLLM, L40S, scale-to-zero) — ~25s warm
- **Frontend:** custom noir **Gradio** app on Hugging Face Spaces (CPU only)
- **Agentic, not a single prompt:** the Inspector works the scene in steps — a full-scene
  **sweep** flags suspects, then it **crops + zooms into each suspect** and re-examines it
  up close to confirm or clear the charge and tighten the evidence box, before filing the
  verdict (plan → act/zoom → verify → synthesize).

## Four powers that set it apart
- ⚖️ **The Trial — two small models argue one case.** A **separate NVIDIA Nemotron model**
  (`Nemotron-Mini-4B`) opens as **THE PROSECUTION**, pressing the filed charges; the guilty
  elements take the stand and defend themselves (Qwen), and the Inspector rules. Qwen sees &
  defends · Nemotron prosecutes · the Inspector judges.
- 🗣️ **Interrogate the Inspector** — post-verdict agentic chat: the same VL model, re-prompted
  with the screenshot + filed case, defends or concedes each charge from the visible evidence.
- 🖼️ **The Reconstruction** — `FLUX.2-klein-4B` (Black Forest Labs) rebuilds the worst guilty
  element *fixed*, live, from the Inspector's own remedy. Before/after on the real pixels.
- 🔊 **The Inspector's Voice** — `Kokoro-82M` local TTS reads the verdict aloud. No API, no keys.

## Badges claimed
- 🎨 **Off-Brand** — fully custom noir/forensic frontend (cinematic intro, evidence desk,
  live investigation, typewritten case file) layered over Gradio.
- 🤖 **Best Agent** — a genuine multi-step visual agent (sweep → zoom-in → verify →
  synthesize) that then **answers follow-up interrogation**, all under the 32B cap.
- 🎬 **Best Demo** — polished app + cinematic trailer + social post.

## Sponsor awards
- 🟢 **Modal — Best Use of Modal** — four GPU backends (vision, FLUX, voice, Nemotron
  prosecutor), all scale-to-zero on Modal, plus case storage on a Modal Volume.
- 🟩 **NVIDIA — Nemotron** — `Nemotron-Mini-4B-Instruct` is THE PROSECUTION: a distinct
  small model arguing the case against the interface, not a decorative text call.

> FLUX.2 Klein (Black Forest Labs) powers the live Reconstruction — credited as the
> model behind that feature, not claimed as a separate prize.

## Also competing for
- 🍄 **Thousand Token Wood** — main-track podium
- 🗳️ **Community Choice**

## Supporting material
- 📓 Field Notes article: https://huggingface.co/blog/kasbsquall/ux-crime-scene
- 📡 Public investigation-traces dataset:
  https://huggingface.co/datasets/build-small-hackathon/ux-crime-scene-traces