| title | AMNESIAC | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| emoji | 🪞 | ||||||||
| colorFrom | gray | ||||||||
| colorTo | red | ||||||||
| sdk | gradio | ||||||||
| sdk_version | 5.50.0 | ||||||||
| python_version | 3.10 | ||||||||
| app_file | app.py | ||||||||
| license | apache-2.0 | ||||||||
| short_description | Reverse-Turing webcam interrogation game. | ||||||||
| header | mini | ||||||||
| fullWidth | true | ||||||||
| tags |
|
"Calibration is complete. Sit forward. The form says human — let's find out if the form is lying." — A.M.N., Human-Authentication Division
Space: https://huggingface.co/spaces/build-small-hackathon/amnesiac
Allow your camera + microphone, press BEGIN CALIBRATION, and talk your way past the machine. (Built for the Hugging Face build-small-hackathon — small models, big nerve.)
📣 Social: see AMNESIAC on X →
👥 Team: @hetanshwaghela
💻 Code (built with OpenAI Codex): github.com/HetanshWaghela/amnesiac
A.M.N. interrogates a live subject — reading the face, the pulse, and the pause before every answer.
For seventy years the Turing test asked one question: can a machine fool a human?
AMNESIAC flips it. You sit in front of A.M.N. — an AI interrogator from the "Human-Authentication Division" — and A.M.N. has to be convinced that you are the human in the room. It doesn't just parse your words. Through your webcam it reads the signals you can't fake:
- 🫀 Your pulse, recovered from micro color-shifts in your skin (remote photoplethysmography — rPPG)
- 😬 Your micro-expressions, from a 478-point facial mesh
- ⏳ The pause before your answer — "The pauses between your words tell me as much as the words."
Every tell feeds a live suspicion meter. Hesitate, and A.M.N. circles back. Crack, and it closes in.
The browser is A.M.N.'s sensory cortex — all face/pulse perception runs on-device, so raw video never leaves your machine; only distilled tells travel over the wire. A warm Modal H100 is the brain + voice: MiniCPM-o reasons about your tells and writes the next question, and a fine-tuned VoxCPM speaks it in A.M.N.'s own cold, clinical register. Media flows over FastRTC / WebRTC so the loop feels live.
| Layer | Technology |
|---|---|
| 🧠 Reasoning + perception brain | MiniCPM-o 4.5 (8B omni) — reads tells, runs the interrogation, writes every line |
| 🗣️ A.M.N.'s voice | VoxCPM 1.5 (750M) + a custom LoRA we fine-tuned on a bespoke 110-clip, 8.7-min voice set |
| 👁️ On-device perception | MediaPipe Face Landmarker (478 pts + blendshapes) and a custom rPPG pulse estimator — 100% in-browser |
| 📡 Realtime media | FastRTC (WebRTC, TURN-relayed) on a FastAPI + Gradio app |
| ⚡ GPU host | Modal (H100, kept warm for the demo) |
| 🎨 Frontend | Hand-built vanilla JS + WebAudio glitch FX — a clinical, Voight-Kampff aesthetic |
Optional protocols turn the screws:
- 🌀 GLITCH — A.M.N. destabilizes. For a few seconds it forgets which of you is the machine, its voice warps and pitch-bends, and the whole frame tears. Then it snaps back, colder.
- 🐠 GOLDFISH — A.M.N.'s memory decays mid-interrogation. It re-asks, forgets, and contradicts itself — and watches whether your story stays consistent when its does not.
A.M.N. doesn't use an off-the-shelf TTS. We curated a clean 110-clip / 8.7-minute dataset in-character, then fine-tuned a LoRA (rank 16) on VoxCPM 1.5. The result is a single, recognizable interrogator — measured, patient, and just slightly wrong — loaded fresh at runtime (320 LoRA tensors, 0 skipped).
- The premise is the gameplay. Not a chatbot with a webcam bolted on — the perception is the interrogation.
- Privacy by design. Face and pulse analysis never leave the browser; only abstract tells are sent.
- Live, not turn-based-feeling. Streaming voice, clause-level synthesis, and a relay-only WebRTC path keep the back-and-forth tight.
- It has a personality you can hear. A voice we trained, a stare that reacts to your biometrics, and failure modes that are genuinely unsettling.
How we built it and what broke along the way (the voice bottleneck, the fine-tuned VoxCPM LoRA, and the realtime plumbing): Read the Field Notes →
Built with too much coffee for the Hugging Face build-small-hackathon.
Now sit forward. Convince it.