EmetEmet

01

The product

Emet is a media authenticity engine. You give it an image (an upload, a URL or a social post link) and it returns AI-generated, edited and authenticity confidence scores. It also returns a verdict, a traffic-light indicator, the reason for the decision and a flag that says when a human should review the result.

02

The problem

Generative images are now hard to tell apart from photographs, and a single classifier score is brittle and hard to trust. People who review content need a confidence they can explain and a signal for when to escalate.

03

My role

Full-Stack & ML Engineer2026

• Designed and built the detection backend in FastAPI, the detection engine and its tests.
• Built a GraphQL service (Django and Strawberry) that wraps the engine, with optional hash-based caching in PostgreSQL.
• Built the Expo mobile client and containerised the stack with Docker Compose.

04

Engineering approach

1. 01A hybrid pipeline runs model inference alongside independent forensic signals (noise residuals, frequency-domain artefacts and edge statistics), then fuses them into calibrated scores.
2. 02Models load once, as a singleton, from either Hugging Face weights or an ONNX export, so the same API can run on a laptop or a small VPS.
3. 03Inference runs asynchronously with timeouts, supports simple batching, and offers a WebSocket endpoint for real-time frame scoring.

05

Architecture

06

Key decisions

07

Technologies

• Python
• FastAPI
• PyTorch
• ONNX Runtime
• Hugging Face
• OpenCV
• WebSockets
• Django
• Strawberry GraphQL
• PostgreSQL
• Docker Compose
• Expo
• React Native

08

Outcome

• ◆Working detection API, GraphQL service and mobile client in active development. Not yet publicly deployed.