Spaces:

build-small-hackathon
/

lost-found-desk

Paused

Requirement	Status	Evidence / action
Every model under 32B parameters	Ready	MiniCPM-V 4.6, MiniCPM5-1B, and Nemotron Embed VL are all below the cap.
Gradio app in official Build Small HF org	Ready after deploy verification	README front matter uses `sdk: gradio`; target Space is `build-small-hackathon/lost-found-desk`.
Demo video	Ready	YouTube is the primary demo link; `demo/lost-found-desk-demo.mp4` remains in the repo as an archived copy.
Social post	Ready	Published on X: https://x.com/JacobLinCool/status/2066147773481951378
Team HF usernames	Ready	README lists @JacobLinCool.
Zero GPU app limit	Ready	This is one ZeroGPU-targeted Space for the user; still verify account-level count before final form submission if multiple entries exist.
README tags	Ready	README contains official tag schema: `track:`, `sponsor:`, `achievement:*`.
README story / tech write-up	Ready	README explains the problem, users, workflow, architecture, deployment, safety rules, and Codex assistance.
Shared agent trace	Ready	Codex-compatible JSONL is published under `traces/` for the official Hugging Face Agent Trace Viewer; public redaction removes local paths, token-shaped strings, and secret-label strings.

Quest and challenge eligibility

Quest / challenge	Eligible?	Requirement	Current status
Backyard AI	Yes	Practical app improving daily life on local/small models.	Primary track; tagged `track:backyard`.
Thousand Token Wood	No	Whimsical / entertainment-first AI-native app.	Not the right framing; this is an operations workflow.
Best MiniCPM Build	Yes	Build with MiniCPM models as a core part of the experience.	MiniCPM-V captions photos; MiniCPM5 handles claim intake, matching, and safe message drafting. Tagged `sponsor:openbmb`.
Best Use of Codex	Yes	Codex-attributed commits in connected GitHub repo or Space history.	This submission commit includes `Co-authored-by: Codex <codex@openai.com>`. Tagged `sponsor:openai`.
Nemotron Hardware Prize	Likely yes	Build with Nemotron models.	Uses `nvidia/llama-nemotron-embed-vl-1b-v2` for retrieval. Tagged `sponsor:nvidia`.
Best Use of Modal	No	Use Modal for development or runtime and note it in README.	Project does not use Modal.
Off the Grid	Yes	No cloud model APIs; local/open model runtime.	No hosted model API dependency. Tagged `achievement:offgrid`.
Well-Tuned	No	Use a fine-tuned model published on Hugging Face.	No fine-tuned model.
Off-Brand	Yes	Custom UI beyond default Gradio.	Svelte UI served through Gradio Server. Tagged `achievement:offbrand`.
Llama Champion	No	Model runs through llama.cpp.	Uses Transformers/Sentence Transformers, not llama.cpp.
Sharing is Caring	Yes	Shared agent trace on the Hub.	Official-format Codex JSONL trace published at https://huggingface.co/datasets/build-small-hackathon/lost-found-desk-codex-traces and grouped in the collection. Tagged `achievement:sharing`.
Field Notes	Yes	Blog/report about what was built and learned.	`docs/article.md`, this notes file, and `ARTICLE.md`. Tagged `achievement:fieldnotes`.
Tiny Titan	Yes by prize-table criteria	Every model is <= 4B parameters.	Documented in README; official submit tag generator does not expose a `tiny` tag.
Best Demo	Ready	Strong app + demo video + social post.	Demo video and published X post are linked from README and notes.
Best Agent	No / weak	Multi-step tool use and planning.	The app has planner-guided intake, but it is not primarily an agentic tool-use app.

Repository hygiene

.env, .env.*, .venv, node_modules, build caches, runtime DB, runtime uploads, and output/ scratch recordings are ignored.
.env.example is kept as a safe configuration template.
Final frontend build in static/ is intentionally committed because the Space should not require a Node build step.
Final demo video is intentionally committed at demo/lost-found-desk-demo.mp4.

Remaining manual items

Open the official submit page, enter lost-found-desk, verify tags, and submit the form.

Final submission summary

Lost & Found Desk is a caption-first lost-and-found return desk for event venues, conferences, gyms, schools, and coworking spaces. Staff photograph one item at a time; MiniCPM-V captions each item; claimants report lost items through a private multilingual assistant; MiniCPM5 and Nemotron Embed help staff narrow candidate matches; and staff confirm every handoff offline. It is a practical Backyard AI app with a custom Svelte UI on Gradio Server, no hosted model API dependency, all models below the 32B cap, Codex-attributed development commits, a public official-format Codex trace dataset for the HF Agent Trace Viewer, and a Hugging Face collection that groups the Space, models, article link, and dataset.