PimEyes

Performs reverse facial recognition by uploading a face image and matching it against a proprietary index of 900+ million publicly crawled images using deep convolutional neural network embeddings. The system extracts facial feature vectors from the query image and performs approximate nearest-neighbor search across indexed face embeddings to identify matching faces across the web, returning ranked results with confidence scores and source URLs.

Implements a multi-stage ranking pipeline that scores facial matches based on embedding distance, facial landmarks alignment, and contextual metadata (image quality, source domain authority). Results are ranked by confidence score (typically 0-100) with visual similarity indicators, allowing users to quickly identify high-confidence matches versus ambiguous results that may be false positives.

Manages user registration, email/password authentication, subscription state, billing information, and account settings. Implements standard security practices: password hashing, session management, two-factor authentication (optional), and account recovery flows. Tracks user search history and removal request submissions within account.

Provides a credit-based search quota system where each facial search query consumes a fixed number of credits (typically 1-5 credits per search depending on subscription tier). Users receive monthly credit allocations tied to subscription level, with the ability to purchase additional credits. The system tracks credit consumption per search and enforces rate limiting to prevent abuse.

For each facial match detected, the system extracts and returns the source URL, page title, domain metadata, and thumbnail preview of the matched image. The system crawls page metadata to provide context about where the image appears (e.g., social media profile, news article, e-commerce listing) without requiring users to manually visit each URL.

Provides an integrated workflow for users to submit removal requests directly to website owners for images containing their face. The system generates templated removal request emails with image details, source URL, and legal basis (GDPR, CCPA, or general privacy concerns), and tracks removal request status. Some integrations with major platforms (social media, search engines) enable automated removal submission.

Implements a multi-tier subscription model (Free, Premium, Professional) with feature gating where higher tiers unlock additional capabilities: monthly search credits, removal request submissions, advanced filtering options, and API access. The system enforces tier-based rate limits and feature availability at the application level.

Handles user-uploaded facial images with preprocessing pipeline: validates file format (JPEG, PNG), detects faces using multi-task cascaded CNN (MTCNN) or similar detector, extracts facial regions, performs quality checks (resolution, blur, lighting), and normalizes images for embedding extraction. Rejects images with no detectable faces or quality issues below threshold.

Extracts deep facial embeddings from uploaded images using pre-trained convolutional neural networks (likely ResNet or similar architecture trained on large facial recognition datasets). Embeddings are 128-512 dimensional vectors capturing facial features. The system maintains a massive index of embeddings from 900+ million crawled images, enabling fast approximate nearest-neighbor search using techniques like LSH (Locality Sensitive Hashing) or FAISS (Facebook AI Similarity Search).

Operates a distributed web crawler that continuously crawls publicly accessible websites, extracts images, and feeds them into the facial embedding pipeline. The crawler respects robots.txt and crawl delays, maintains freshness of indexed images, and deduplicates images to avoid redundant indexing. Crawled images are processed asynchronously to extract facial embeddings and add to the main index.

Generates comprehensive privacy audit reports summarizing where a user's face appears across the internet, including statistics on number of matches, top domains hosting images, geographic distribution of sources, and timeline of when images were first indexed. Reports can be exported as PDF or CSV for record-keeping or legal purposes.