What I Built

Nose ID — a biometric dog identification system. Instead of a collar
tag, chip scanner, or manual lookup, it identifies a dog the way security
systems identify people: from a unique physical trait. In this case, the
dog's nose-print — the ridge and texture pattern on a dog's nose, which is
as individual to each dog as a fingerprint is to a human.

Scan a dog's nose → the system tells you who it is (with a spoken voice
response) or offers to enroll it if it's new.

Demo

How It Works

1. Embedding extraction (the core ML piece)
Rather than training a nose-print classifier from scratch, I used a
pretrained MobileNetV2 (ImageNet weights) as a feature extractor,
stripping its classification head and using the penultimate layer's
pooled output as a 1280-dimensional embedding vector. This is a standard
transfer-learning technique — mid-level CNN filters trained on general
images still respond well to texture, edges, and ridge patterns, which is
exactly what distinguishes one nose-print from another.

2. Matching
Each enrolled dog's embedding is L2-normalized and stored. A new scan's
embedding is compared against all stored ones using cosine similarity.
A similarity above a tuned threshold (0.90) counts as a match.

3. AI-generated personality
On enrollment, Google Gemini generates a short, fun bio for each dog
from its name and breed.

4. Voice confirmation
On identification, ElevenLabs turns the result into a spoken response
— "Welcome back, Bella. Access granted." or "I don't recognize this nose.
Would you like to enroll?"

Tech Stack

• Frontend: React (Vite)
• Backend: FastAPI (Python)
• ML: PyTorch, pretrained MobileNetV2, cosine similarity matching
• Google Gemini — dog bio generation
• ElevenLabs — text-to-speech voice confirmation

Challenges I Ran Into

Both Gemini and ElevenLabs changed parts of their APIs the same week I
built this:

• Google's new API key format required switching from the legacy generateContent endpoint to their new Interactions API.
• ElevenLabs' free tier restricts which voices are callable via the API, even though the full library shows in their dashboard — had to query /v1/voices to find one my account could actually use.

Try It

Prize Categories

• Best use of Google AI
• Best use of ElevenLabs

weekendchallenge