Face Search, Done Responsibly
We’re an OSINT orchestrator and search engine. We index and analyze only publicly available information.
Every search is targeted and tied to a specific case. Our approach combines knowledge signals, reverse
image search, non-biometric facial analysis, and Perceptual Lookalike Ranking to surface relevant results
while being compliant.
We’re driven by both compliance and innovation. Where the law permits, we may
use facial recognition. In other cases, we rely on non-biometric methods.
Search: Two complementary data retrieval lanes
Our system runs two complementary retrieval lanes in parallel, then applies privacy-preserving matching and curation to deliver confident, reviewable results.
1. Image-Led Retrieval
Reverse image search, our index
We search our internal image database of hashes and thumbnails to find exact or near-duplicate matches of the uploaded photo. Many of these images are already linked to public posts and pages, which often serve as strong starting points for Knowledge Search. A direct reverse image search alone can sometimes surface a social profile without any need for facial analysis.Reverse image search, third party
We also federate searches to major image engines and tools like Google and Yandex. These help find duplicates and visually similar photos that may show the same person in different contexts. The returned candidates can be filtered using non-biometric facial analysis with ranking, or in limited compliant cases, by ephemeral facial recognition.Nonbiometric facial database, our index
We search for faces based on general, non-unique facial traits like hair color, eye color, presence of glasses, rough age group and other traits. This method returns lookalikes, not specific identities. It often surfaces adjacent candidates we can later verify using knowledge-based signals.Vectorized facial characteristics database, used sparingly
We maintain a small, restricted database of vectorized facial characteristics. These are not face landmarks and cannot be used to reconstruct the original image. This database is used only in two cases: (a) in jurisdictions where opt-in consent is not legally required and use is otherwise compliant, or (b) where we have explicit, revocable consent from the data subject. Data is retained for two years. In practice, most searches are completed without accessing this database.2. Knowledge-Led Retrieval
Knowledge Search across names, aliases, locations, emails, and handles
We search our open-source index, along with third-party engines and services, to find public social profiles, mentions, media articles, and other artifacts connected to the individual and their images.Nickname and handle synthesis
We generate likely usernames based on names, initials, and common patterns to locate social accounts and mentions that may not match exact inputs.Advanced query techniques
We apply broad Google dorking style queries across multiple engines to surface indirect public pages, not only Google.Cross-lane feedback
When knowledge results include images, they are looped back into Image-Led Retrieval. We run reverse image searches and, when appropriate, apply ephemeral matching to strengthen the evidence chain. We also perform visual distortion correction and face alignment. In some cases, AI is used to slightly modify the image to either expand or narrow the search and matching scopes.Surfface currently limits person-related information to social media profiles and image sources. We prioritize legal and ethical standards.
First we ensure compliance, then we release.
Matching: Two face approaches
1. Nonbiometric Facial Analysis and Perceptual Lookalike Ranking
This is facial analysis, not facial recognition. It intentionally avoids scanning face geometry, faceprints, landmarks, or embedding vectors. There’s no enrollment, no templates, and no persistent face identifiers.
How It Works
Ingest and detect
We receive an image or video frame and detect face regions.Attribute extraction only
Inside each detected face box, lightweight classifiers extract general, non-unique attributes. These include hair color, eye color, presence of glasses or facial hair, coarse age group,rough face, nose shape and other traits. We do not create or store faceprints, embeddings, or facial geometry vectors.Search by filtering, not recognition. Lookalikes by design
Each query is turned into a set of attributes. We retrieve candidates who share those traits and rank them based on overlap and tag weight. There’s no identity matching—just filtering. Results are intentionally broad. Multiple people may score equally, and that’s expected. This level of matching is often enough to uncover useful images from pools built through Knowledge-Led Retrieval and third-party reverse image search.Perceptual Lookalike Ranking
To make results feel more natural and accurate we built Perceptual Lookalike Ranking. It’s our proprietary neural model that, simply speaking, mimics how people judge facial similarity, but without using biometric geometry. The model heavily relies on Nonbiometric Facial Analysis results.
It began as an effort to improve ranking and search accuracy using biometrics, along with experiments with novel approaches like A View From Somewhere: Human-Centric Face Representations. Over time, our developments and tests evolved into a bigger idea: ranking could work without biometrics at all. It took over a year to get it production-ready, but we did it. The result is a practical similarity score that feels human, but stays fully outside the boundaries of biometric identification.
2. Ephemeral Face Recognition Matching, When Permitted
In rare cases, a small set of candidates remains ambiguous, and non-biometric methods aren’t enough to confidently resolve the match. If—and only if—the legal jurisdiction and case context allow, we may apply face recognition as a narrow filter to deduplicate or cluster these remaining candidates.
Most searches are resolved without it, thanks to Knowledge Search, Perceptual Lookalike Ranking, and our Context-Aligned Similarity AI.
Curation Engine: Context-Aligned Similarity AI
Because we use facial recognition only in rare, limited cases, we put a strong focus on data fusion and context.
Our system analyzes non-image signals that travel with photos such as captions, timestamps, locations, poster profiles, nearby names and handles, and more. Behind the scenes, we build an ephemeral graph of connections for the current case, then apply Context-Aligned Similarity AI to:
- tighten the candidate pool by prioritizing results that match the context of the query
- down-rank or remove results that likely refer to different people
- re-rank the list so similarity reflects both visual resemblance and supporting evidence
This rigorous approach significantly slows processing and increases server costs, but it ensures compliance. We plan to introduce token-based monetization, allowing users to allocate more resources for deeper and higher-quality results. If regulations require it, we will use tokens not to offset server costs, but to fund manual face recognition by staff in India while continuing to deliver results.
Compliance by Design
We take trust seriously. We build our system to avoid getting it wrong.
You may not use our service or the information it provides to make decisions about consumer credit, employment, insurance, tenant screening, or any other purpose that would require FCRA compliance. Surfface does not provide consumer reports and is not a consumer reporting agency. (These terms have special meanings under the Fair Credit Reporting Act, 15 USC 1681 et seq., (“FCRA”), which are incorporated herein by reference.) Surfface does not make any representation or warranty about the accuracy of the information available through our website or about the character or integrity of the person about whom you inquire.