For brands, retailers and creators
Your customer applied it. Now help them get the result.
Try-on tools predict a result; siOsi reads a photo of the real one. Fifteen technical labs covering makeup application and face geometry, plus a full 12-season color analysis. It is live, in nine languages, and you can run an analysis yourself right now - free, no account needed - before you talk to anyone here.
Where the returns actually come from
Color tooling in beauty tends to fail in the same three places: the camera distorts skin tone before any model sees it, filters simulate a result instead of analysing one, and the shade ranges brands stock skew light, underserving deeper complexions.
60%+
More than 60% of online beauty shoppers have decided not to purchase a beauty or cosmetic item online because they didn't know what color or shade to choose.
Google, Inspired Shopping Report (7 April 2022)41%
Among online beauty shoppers, 41% have returned an item because it was the wrong shade.
Google, Inspired Shopping Report (7 April 2022)6.19%
Beauty has its own blended return rate, distinct from other retail categories: Appriss Retail's analysis of 40,000 stores found Beauty returned at 6.19%. For context, NRF separately surveyed a broader set of retailers - including categories like grocery stores that aren't part of this category breakdown - and found a 10.0% median across that wider group, so the two figures are not computed on the same basis.
Appriss Retail & National Retail Federation, "2018 Consumer Returns in the Retail Industry" (December 2018)For scale, across all US retail: $849.9B (15.8% of sales). National Retail Federation & Happy Returns (a UPS company), 2025 Retail Returns Landscape
What you get
Color Analysis API
A photo in, and back out: 12-season classification, undertone, palette hex values, and a confidence score for the read.
Makeup Analysis API
Fifteen labs scored per photo, returned as structured per-lab results with written feedback rather than a single number.
White-label embed
The analyser running under your brand, on your domain, with your type and colors. Custom development, not an off-the-shelf feature.
Palette to SKU matching
Analysis output mapped onto your catalogue, so a season becomes a shortlist of products you actually stock. Custom development, not an off-the-shelf feature.
Who this fits
siOsi fits five shapes of business. Where product is sold, the value shows up as fewer wrong-color purchases and returns; where it is not, the case is consistency across staff or engagement with an audience, and each entry says which one applies.
Beauty e-commerce and marketplaces
The strongest fit. Wrong-color purchases are the failure a photo-based read at checkout or in a size-and-shade guide addresses directly. Whether that moves your returns line is what a pilot exists to measure, not a number this page will claim.
Cosmetics brands selling direct
Personalisation on your own site using the Color Analysis API, with the same returns logic as e-commerce generally. Matching the read onto your own catalogue is custom development, not an off-the-shelf feature.
Multi-brand retail
One consistent read across every brand on the floor, plus a tool a floor advisor can hand a customer instead of eyeballing a shade. Returns economics apply here too, at the retailer level rather than per brand.
Salon networks and chains
One integration across many locations, so every stylist gives the same color read instead of each improvising. The case here is consistency and training rather than returns, which makes it a fit for multi-site groups rather than a single chair.
Beauty creators and influencers
An audience tool under your own name, needing no integration beyond a subdomain. Because the recommendation catalogue already carries affiliate links, the commercial shape here can be revenue share rather than a flat fee. There is no product being returned, so the value case is engagement and affiliate revenue rather than returns.
What the engine actually returns
Everything above describes the analysis in prose. Below is what the same engine actually returns, derived from the real schemas rather than written up for the page: color analysis and makeup analysis, the two Live APIs above, plus the myth-busting and product-recommendation content the same engine generates from that analysis.
A live session a buyer can open right now - full color and makeup labs, face geometry, and confidence scores, no login required. Open a real Deep Autumn read
- A twelve-season color classification for the photo
- Undertone: warm, cool, neutral, or olive
- A recommended palette and an avoid palette, as hex values
- A confidence score attached to every read
- Fifteen labs in total, per photo: twelve on makeup application, three on face geometry
- Skin-hair harmony, contrast, and eyebrows analysed too, separately from the fifteen labs above
Makeup application (12)
- Undertone
- Flashback
- Coverage
- Pores
- Texture
- Transfer
- Transitions
- Longevity
- Oxidation
- Shimmer
- Blending
- Creasing
Face geometry (3)
- Lip shape
- Face shape
- Eyelid shape
Also analysed
- Skin-hair harmony
- Contrast
- Eyebrows
Analysis, not a filter
A filter renders a plausible image. An analysis returns a classification you can act on, with a confidence score attached, so you know when the read is weak and can ask for a better photo. That difference matters most where the industry does worst: shade ranges themselves are not evenly distributed across skin tones, which is a documented assortment gap rather than a matter of taste.
A peer-reviewed content analysis of 49 mainstream cosmetics brands found their foundation shade ranges are not spread evenly across skin tones: brands consistently stock far more shades for light-to-medium skin than for darker skin, and darker shades are more likely to be available through e-commerce than in physical variety or department stores. Cynthia M. Frisby, "Black and Beautiful," Advances in Journalism and Communication, Vol. 7 No. 2 (2019)
- Fifteen labs per analysis: twelve on makeup application, three on face geometry
- Twelve-season color classification
- Nine supported languages
Uptime is published publicly at status.siosi.me
What a photo can and cannot prove
Different checks need different evidence, and a tool that pretends otherwise is guessing under your logo. This is the standard a pilot is scoped against: the checks we turn on are the ones your capture flow can actually support, and a weak read gets flagged as weak rather than dressed up as certainty.
Coverage, blending, transitions, creasing, symmetry
One well-lit photo, taken after application.
Flashback
A photo taken with flash. Ambient light cannot show it.
Oxidation and longevity
A fresh photo plus a later one under comparable light.
Transfer
Physical evidence: a blot test, a collar, or the customer telling you.
A measured foundation shade
Out of scope. That needs calibrated color capture, and the FAQ below says so plainly.
How a pilot works
01
Scoping call
Twenty minutes on what you sell, where the analyser would sit, and which single metric would make the pilot a success.
02
Sandbox access
Access to the endpoints or a branded embed against test traffic, so your team can judge output quality on your own photos.
03
Thirty-day pilot
A fixed window measuring the metric agreed in step one. If it does not move, we both learn something cheap.
Pilots are paid, from $1,500 for the thirty-day window, and the exact figure is fixed on the scoping call against scope and volume. Paid is deliberate: it keeps both sides honest about whether this is worth doing.
Questions brands ask
What happens to the photos we send?
Photos are transmitted over HTTPS, processed to produce the analysis, and are not sold or used to train models. Retention windows and deletion are set per agreement, and the current consumer policy is published in full at /en/privacy.
How is accuracy reported?
Every analysis returns a confidence score alongside its result, so a weak read is visible as a weak read rather than presented as certainty. Low-confidence photos are flagged so your interface can ask for a retake instead of guessing.
What does siOsi not do?
To be exact about scope: siOsi does recommend products, but not by measurement. It analyses makeup application and selects from siOsi's own curated catalogue on that basis. What it does not do is read skin reflectance and compute a color distance against a foundation range, so it will never tell a customer they are a 3.5N. If a measured shade match is what you need, that is a different product and we will say so on the call.
How much integration work is it?
The APIs take a photo and return JSON, so the usual shape is one server-side call and a rendering decision on your end. The white-label embed needs no integration beyond a subdomain and your brand assets.
What does a pilot cost?
Pilots start at $1,500 for the thirty-day window. The exact figure is set on the scoping call, against the scope and volume agreed there, because a thirty-day embed test and a high-volume API integration are not the same commitment.
Who owns the analysis output?
You do, for anything generated from your traffic, and that is written into the pilot agreement. siOsi retains rights to the underlying models and palettes.
Start a conversation
Tell us what you sell and which piece you are interested in. Replies come from a person, usually within two working days.
Who you would be talking to
Bárbara Sia
CBO, Co-Founder
Demetrio Koro
CTO, Co-Founder