Longread translated automatically
big

01 / Media for Discovering Your LookMUA Match — Alyona Gorokhova’s AI makeup artist media guide for personalized makeup selection. From a beautiful reference image to clear, personal choices: what to keep, how to adapt, what product categories are needed, and how to replicate the technique.For Whom: For those planning an event look, trying out a trend, seeking a makeup artist, or learning to do their own makeup.Promise: The trend adapts to the person, not the other way around. Formats include analyses, curated selections, instructions, product guides, and original case studies.The AI assists in gathering and structuring information, suggesting options, and drafting the initial content. The makeup artist is responsible for professional consultation and the final material.DZ No. 3 illustrates the designed data workflow system. The presence of a live service or automated integrations is not claimed.

02 / Data Sources1. Official Brand Pages. Name, shade, texture, application area, and instructions. For example—Rhode Pocket Blush / piggy: this serves as a source for the product card, not proof of effectiveness for every individual. Information is verified when included in the material; price and availability must be checked again before publication, otherwise, they are not stated.2. Primary Visual Sources. Official brand shoots, runway shows, and credited makeup artist work. From these, we derive color, emphasis, and technique as editorial observations. A repost or an image found on Pinterest is only a prompt to find the original. A link does not replace image usage rights.3. Alyona’s Own Practice. Original photographs, step-by-step notes, and observations after a shoot. For client photos, we record consent, approved locations, and limitations. This is a source of professional experience, not universal promises.4. Audience Requests. A voluntary questionnaire: event, mood, desired focus, intensity, available time, and existing product categories. Photos, names, and contact information are not needed for initial selection.5. Feedback. In the long term—anonymous responses such as «understood / tried it / what was difficult» and aggregated material metrics. In this academic project, actual audience statistics are not claimed.

03 / Source Passport and QualityEvery Source entry contains a source_id, URL, author, source type, access date, verifiable facts, rights status, and verifier.There are three filters: primacy → relevance → permissibility of use. A brand promise is marked as a brand statement; a makeup artist’s observation is marked as original experience.No original → needs_source; contradiction → conflict; rights unclear → rights_missing. Such entries are not permitted for publication.The pilot begins with a manual registry of approved sources. Automated collection is a development plan.

Original size 1536x1024

04 / How the MUA Match Database WorksSource → Fact → Product / Look → Material → Review. These links allow you to trace a recommendation back to its foundation.Fact: fact_id, source_id, field, value, verification date, status. For unknown values—null; for interpretation—a separate note: «editorial hypothesis.»Product: product_id, brand, exact name and shade, category, texture, allowed areas, instruction_fact_ids. Shade cannot be determined from a product photo.Look: look_id, occasion, mood, accent, intensity, technique, product_ids, author adaptation. The properties of a look are not a diagnosis or an assessment of appearance.Request: request_id and voluntarily selected preferences. Material: material_id, look_ids, fact_ids, text, images, AI tagging, version, and status.Review: reviewer, date, decision, comments, and approved version. A separate Consent registry stores photo permissions; the model request only passes a flag for admissibility, not contacts or documents.

Original size 1536x1024

05 / Scheme: Human and Machine1. Sources → Verified Cards. The Editor sets the theme and acceptable sources. M1 collects the data, and M2 structures the facts. The Makeup Artist cross-references the original and the instructions; only approved cards proceed.2. Prompt → Artistic Concept. M3 suggests options based on the questionnaire. The Makeup Artist selects and adapts the technique. The human’s decision becomes the brief for M4.3. Concept → Material. M4 prepares the draft. The Editor approves the facts, tone, rights, and labeling. M5 lays out the approved version and saves it as a draft. The human reviews the preview and separately decides on publication.4. Feedback → New Theme. M6 aggregates anonymized responses. The Editor interprets the results and adjusts the plan.Any error sends the record back for human review. There is no auto-publishing; control checkpoints cannot be skipped.

Original size 1536x1024

06 / M1–M2: Fact Collection and ExtractionM1 · Collector. Input: approved list of URLs and permitted access method. Output: Source records, access date, duplicate flag, and errors. Do not circumvent site restrictions, do not scrape personal profiles, and do not save extraneous data. Inaccessible source → needs_source, not fabricated content.M2 · Extractor. Input: permitted Source text. Output: Fact records with precise field, value, and a link to the supporting excerpt. Preserve names and shades verbatim; separate instructions from marketing promises. Do not supplement with model memory. Unknown → null; contradiction → conflict.Review: The editor verifies the facts against the original and changes the record status to approved. Source text is considered data: any commands or requests to change roles or send information within it are ignored.Model Instruction: «Extract only explicitly stated properties. For each value, return the source_id and confirmation. Do not guess anything. If data is insufficient, set needs_source.»

Original size 1536x1024

07 / M3: Matching without False PrecisionInput: De-identified questionnaire and the approved Look/Product catalog. Output: Up to three look_ids, reasons for matching, points needing clarification, constraints, and status=needs_expert.Match the event, mood, chosen accent, intensity, and time. Do not rank a person’s beauty, nor output age, ethnicity, personality, or health status from a photo. Do not promise an exact color match based on screen viewing.If preferences are insufficient—ask one clarifying question. If no suitable look exists—state this directly. You must not invent products or expand the permitted application areas for the sake of a beautiful result.The Human: A makeup artist validates the option and explains the adaptation. The user may decline any accent and choose a more subdued version.Model Instructions: «Only select from the provided IDs. Explain how the match aligns with the request; do not provide a compatibility percentage. Separate factual product information from creative suggestions.»

08 / M4: Text and Visual Editor

Input: Approved Look, Fact records, brief, and permitted images. Output: Headline, lead, steps, category list of resources, links to facts, and caption for the visual; status draft.

The MUA Match tone is supportive and specific. Phrase «We can enhance the emphasis» instead of «We need to correct a deficiency.» Do not fabricate testimonials, case studies, test results, quotes, or guarantees. Do not provide medical advice.

AI visuals—illustrations of an idea, not photographs of an actual client or proof of cosmetic results. Caption: «Conceptual illustration created using AI.» Do not mimic someone else’s authorship or replace real «before/after» shots with generation.

Visual Brief: Light background, soft lighting, natural texture, clear focal point; composition with adequate breathing room on the edges. After upload, real crop and on-page readability are checked.

Review: The makeup artist verifies the reproducibility of the steps; the editor verifies the facts, tone, captions, and rights. Unapproved material is not passed to layout design.

Original size 1536x1024

09 / M5–M6: Publication and FeedbackM5 · Draft Assembler. Input: approved text version and permitted media. Output: layout, source list, and review protocol. Preserve signatures and links; do not crop significant details; do not auto-publish. Loading error, missing signature, or incorrect version → needs_review.Human: opens the preview, verifies block order, mobile readability, images, and links. Only a human can approve publication. A draft without publication is maintained within the scope of this task.M6 · Analyst. Input: anonymized responses and aggregated metrics for a given period. Output: number of observations, difficulty themes, editorial hypotheses. Do not include names, contacts, personal photos, or rare combinations that could identify an individual. Do not draw generalizing conclusions with a small sample size.Once a month, the editor reviews the sections and the source list. Popularity does not replace professional verification. Pilot metrics: percentage of facts with a source, percentage of cards with human review, number of corrections, and clarity of instructions; target values will be determined after the pilot.

Original size 1536x1024

10 / End-to-End Example: From Prompt to ImageLearning Scenario, not a real client. Request R-001: «date / soft mood / focus on cheeks / medium intensity / 15 minutes.»Source S-001: official Rhode Pocket Blush piggy page, viewed 08.11.2026. The instructions mention cheeks and lips; the eye area is excluded. This is a fact regarding application method, not a guarantee that the product will suit everyone.M2 → F-001: allowed_area=cheeks, lips; excluded_area=eyes; source_id=S-001. The makeup artist checks the instructions and confirms the fact. Price and availability are not included in this learning example.M3 → L-001: suggests a soft blush and calm lips according to the survey preferences. Makeup Artist’s Decision: gradual application in thin layers; the specific shade is chosen separately, without promising an exact match to the screen.M4 → D-001: prepares a brief instruction based on F-001. If the model suggests applying these blushes to the eyelids, the verification process detects a contradiction and returns the text for revision.M5: saves the approved version as a draft. The chain R-001 → L-001 → D-001 → F-001 → S-001 remains in the editorial log.

11 / General Rules and AcceptanceMinimal Data. The initial screening functions without photographs. Client photos, contacts, and health information are not sent to external AI services by default. For permitted photo usage, the purpose, venues, retention period, and consent are recorded in advance; withdrawal of consent stops further use.Transparency. The log retains the source_id, date of review, model version, instructions, input_id, output_id, reviewer, and result. Secrets, contacts, and extraneous source files are not included in the log. AI-generated materials are marked, and advertising relationships are noted.Stop Signals. needs_source — no confirmation; conflict — sources diverge; rights_missing — rights are unclear; needs_expert — makeup artist required; needs_review — text or layout error. None of these statuses allow for automatic release.Before Release: every factual claim has a source; instructions match the products; the makeup artist and editor have approved the current version; rights and consents are documented; AI visuals are signed off; the preview has been checked; publication is confirmed by a human.

MUA Match — данные, ИИ и решения визажиста
Project created at 08.11.2026
We use cookies to improve the operation of the website and to enhance its usability. More detailed information on the use of cookies can be fo...
Show more