Walk through a passport e-gate at Heathrow on a Tuesday morning. The camera blinks, the glass slides open, you keep moving. The software that just decided you are who you say you are - that compared a live face to a stored photo without flinching at your jet lag, your new glasses, the bad light - was likely trained by a 61-person company in San Francisco's Presidio. The company is called Paravision. You weren't supposed to notice it. That's the point.
Who they are now
A quiet company in a loud category.
Paravision sells biometric software. Face recognition, liveness detection, deepfake detection, age estimation, the contactless corridor at the airport. The product list reads like the table of contents for a privacy debate, and yet Paravision is not the company you've seen on cable news. It does not run a consumer app. It does not scrape the open web. It does not, as CEO Doug Aley has been at pains to say in interviews, sell to the markets it considers ethically risky. It licenses SDKs to enterprises and governments, charges by deployment, and goes home.
The result is a company with a strange shape. Roughly 61 employees. About $6.7M in estimated revenue. $47M raised over the years. And a customer list that includes a Five Eyes government partner, a handful of airports, banks running remote identity verification, retailers experimenting with pay-by-face, and stadium operators tired of ticket fraud. Wide reach, narrow team.
The problem they saw
Faces are the worst password.
Faces are also the most convenient one. That tension is the entire industry. A face is easy to capture, hard to forget, and - thanks to a decade of generative models - increasingly easy to fake. The same algorithms that let a bank verify a customer at 11pm from a kitchen counter can be steered, in the wrong hands, into the kind of surveillance grid no democracy wants to admit it's building. The same model that opens a phone can be embarrassed by a wig.
Paravision's wager is that this isn't a reason to abandon biometrics. It's a reason to build them properly. Properly meaning: tested across demographics, audited by the U.S. government's National Institute of Standards and Technology, deployable on the edge so data doesn't have to leave the device, and packaged in a way enterprises can integrate without writing a research paper.
The founders' bet
A consumer photo app, an FTC notice, and a pivot.
Paravision did not begin as Paravision. It began, in 2013, as Everalbum - a consumer photo storage app that wanted to be the Dropbox for memories. The app accumulated users, accumulated faces, and eventually accumulated regulatory attention. In 2018, the team made a decision that, in retrospect, looks almost stubbornly principled: shut down the consumer side and become an enterprise face recognition company instead. Sell the technology, not the photographs.
The pivot was bumpy - companies that stop being a thing rarely look graceful while doing so - but it sharpened the mission. In August 2019, the company rebranded again, this time to Paravision. In the same week, its algorithm placed first in NIST's Face Recognition Vendor Test 1:1 evaluation. A new name, a new pitch, and a benchmark that mattered, all within a few days. Industries don't usually punctuate themselves that neatly.
Doug Aley became the public face of the company. Benji Hutchinson, who had spent the previous twenty years inside the world's largest multimodal biometric programs, joined as President and COO in 2021. The leadership cadence has stayed quiet since: a CFO, a CTO, a Chief Product Officer, an SVP for Business Development, a General Counsel who also runs data acquisition. Smaller than its competitors. Older than its peers. More careful than either.
// Paravision - A Field-Sketched Timeline
The product
Five SDKs, one stubborn obsession.
The Paravision catalogue is small enough to fit on a single slide and big enough to cover the awkward middle of most identity workflows. Face Recognition matches one face to another, or one face to many. Liveness Detection checks that the thing in front of the camera is a human being and not a video, a mask, or a phone held in front of another phone. Deepfake Detection looks at the image itself and asks whether it was generated or manipulated. Age Estimation infers age from a frame, which sounds simple until you remember every regulator in Europe currently has opinions about who is allowed to do what online. The Contactless Corridor stitches several of these together for high-throughput physical access, which is what airports and stadiums actually want.
All of it ships as software development kits - libraries other engineers integrate. Edge, cloud, server, mobile. Nothing is dependent on Paravision running a giant data lake on your behalf. The pitch is intentionally unglamorous: better than the open source models, more accurate than the giants, more honest about demographic performance than either.
The proof
NIST, the Five Eyes, and a Luminary tag.
For most B2B companies, the proof point is the logo wall. Paravision has one of those - PopID for face payments, Toppan for identity in Asia, and the kind of unnamed government clients you have to read between the lines to identify - but its real currency is the benchmark. NIST has been measuring face recognition vendor performance for years, and Paravision returns to its leaderboards with the regularity of a runner who actually trains. In 2019 it placed first in the 1:1 evaluation. It has continued to rank near the top across configurations since.
The deepfake work is newer but has its own receipts. The first version of Paravision Deepfake Detection emerged from two years of R&D in collaboration with a Five Eyes government partner, trained on a proprietary dataset of more than a million consented images. In October 2024 that partner moved the program out of research and into production. In June 2025 the company shipped Deepfake Detection 2.0, with up to 97% improvement in error rates and the ability to spot fully synthetic faces from diffusion and GAN models - not just swapped or expression-edited ones. The same year, analysts named Paravision a Luminary in the Deepfake & Synthetic Identity Prism report.
The mission
Convenient. Accurate. Ethical. Pick three.
Most companies in this category will tell you they care about ethics. Paravision puts it in the org chart. Its general counsel also runs data acquisition, which is a small organizational detail with a large philosophical implication: the lawyer signs off before the dataset gets used. The company publishes demographic performance, refuses certain markets, and frames the work as identity authentication rather than identification. The difference between those two words is the whole industry, and Paravision picks the smaller one on purpose.
It is, of course, easy to talk about responsible AI when you don't have a billion-user product to feed. Paravision's restraint is partly principle and partly business model. By selling SDKs to vetted enterprises and governments, the company never has to defend a consumer feed full of someone else's faces. It is a B2B company in a category that desperately needs a few more of those.
Why it matters tomorrow
The synthetic face problem is just starting.
In 2024, generative models started producing faces a casual viewer could not reliably distinguish from a photograph. In 2025, they started producing video. The 2026 fraud reports - the ones banks share with each other but not with you - all describe a sharp rise in deepfake-driven account takeovers and synthetic identity fraud. None of this is theoretical anymore. Onboarding flows, payments, age-gated commerce, government benefits: every workflow that asks a human to prove who they are now also has to ask whether the human in front of the camera is, in fact, a human.
Paravision's answer to that is unglamorous. Better models. More representative data. Detectors that update faster than the generators. A refusal to sell into the parts of the market that would weaponize the same tools. And SDKs that run at the edge so the verification can happen on the device, not in someone else's cloud. The bet is that, as identity fraud becomes the dominant fraud, the boring infrastructure company will be the one that quietly handles a meaningful share of it.
Back at the e-gate.
The Tuesday morning at Heathrow ends without incident. The glass slides, you keep walking, you make your connection. Somewhere upstream of that gate, in a server you will never see, a Paravision model returned a match score in milliseconds, a liveness check confirmed your face wasn't a printout, and a deepfake detector quietly verified that the photograph stored against your record had not, in fact, been synthesized by anyone. The company that built that pipeline doesn't want a thank-you. It would prefer the second-best outcome: that you keep not noticing.