README.md

metadata

title: EYEWITNESS
emoji: 👁️
colorFrom: yellow
colorTo: red
sdk: gradio
sdk_version: 6.17.3
app_file: app.py
pinned: false
license: apache-2.0
tags:
  - build-small-hackathon
  - thousand-token-wood
  - openbmb
  - minicpm
  - modal
  - tiny-titan
  - off-brand
  - off-the-grid
  - llama-champion
  - game
short_description: You saw the thief for 3s. Can your memory convict him?

👁️ EYEWITNESS

You saw the thief for 3 seconds. Your memory is the only witness.

A street camera catches a 3-second glimpse of a thief. You describe them from memory to a police sketch artist — then you must pick the real culprit out of a lineup built from your own mistakes. The worse your memory, the harder the lineup.

Four escalating ranks: shorter glimpses, bigger lineups, and from INSPECTOR up, the culprit changes their look between the crime and the lineup ("he's been to a barber since").

How the AI is load-bearing

Stage	Who does it
Spoken testimony → verbatim text (EN/ES)	Cohere Transcribe 2B — a court reporter, not an interpreter: it writes down what you SAID
Your messy testimony → strict attribute JSON	MiniCPM5-1B fine-tune (LoRA, slot-filling over a closed vocabulary — exactly what a 1B is great at)
The composite sketch, the lineup, the scoring	A deterministic engine. Ground truth is authored, never detected — the game always knows the culprit's exact face, so difficulty is honest and the scoring is transparent
The culprit's verdict taunt — a joke about the CRIME and his fate	MiniCPM5-1B (base; improvises live, best-of-6 + validators, with an authored per-crime floor)
The culprit's voice, cloned live	VoxCPM2 (anchor voices cast by sex/age, selected by measured pitch)
Crime flavor + case bank	MiniCPM5-1B-GGUF batch-generated via llama.cpp on Modal

Runtime stack ≈ 5.7B parameters, all inside this Space. No cloud APIs. Every reveal shows what you said → what the artist drew → the truth, side by side — fairness by transparency.

Team

• HF: Fcabla

Links

• Demo video: https://youtu.be/BKOjBMKcb_E
• Social post: https://x.com/i/status/2065901377465819231
• Source code: github.com/fcabla/eyewitness
• Fine-tuned parser: Fcabla/MiniCPM5-1B-eyewitness

Built for the Build Small Hackathon (Thousand Token Wood track) with Claude Code doing the engineering. Lineage: every eyewitness-memory study ever published × police procedural fantasy — the game is honest about how bad your memory is, because we authored the ground truth.