How We See Beauty in Gifts

gifts.supply is an AI-powered analytics platform for Telegram Gift NFTs. We don't just track rarity โ€” we measure visual harmony, semantic meaning, and color science to understand what makes a gift truly special.

6
AI Models
~4.8B
Parameters
400+
Languages
3
Search Layers
35+
Achievements
42
Services

Scoring Philosophy

Most NFT platforms rank items by rarity alone. We believe visual harmony matters more for collectibles โ€” that's why aesthetic score carries 62% of the final Collector Score.

collector_score = 0.23 ร— rarity + 0.62 ร— aesthetic + 0.15 ร— serial
23%
Rarity
How rare are the attributes
62%
Aesthetic
Visual harmony of the combo
15%
Serial
Number beauty & rank

Rarity is weighted: 65% model + 15% backdrop + 20% symbol. Serial score rewards both low numbers (rank 30%) and beautiful patterns (beauty 70%) โ€” palindromes, meme numbers (69, 420, 1337), repeating digits.

Compatibility Engine

The aesthetic score measures how well a gift's components fit together. Three compatibility pipelines run across every possible attribute pair:

aesthetic = 0.56 ร— symbolโ†”model + 0.18 ร— backdropโ†”model + 0.26 ร— collectionโ†”symbol

Symbol โ†” Model Compatibility (56%)

The primary compatibility โ€” does the symbol make sense with the model? Four AI dimensions are blended:

symbol_model = 0.45 ร— semantic + 0.28 ร— visual + 0.10 ร— strict + 0.17 ร— cross
  • Semantic (45%) โ€” Text embedding similarity via BGE-M3 (1024d). Do they mean similar things?
  • Visual (28%) โ€” Image embedding similarity via DINOv2 ViT-L/14 (1024d). Do they look alike?
  • Strict (10%) โ€” Multiplicative: semantic ร— visual. High only when both agree.
  • Cross-modal (17%) โ€” SigLIP (1152d): can the AI recognize the model from the symbol's image, and vice versa?

Backdrop โ†” Model Compatibility (18%)

Backdrops are radial gradients โ€” no meaningful image embeddings. We compare them semantically only (BGE-M3 text embeddings). Separately, we compute color separation in dual color space (70% HSV + 30% OKLCH) for the monochrome/contrast achievement system.

Collection โ†” Symbol Compatibility (26%)

Does this symbol fit the collection's theme? Semantic-heavy:

coll_symbol = 0.60 ร— semantic + 0.20 ร— visual + 0.15 ร— strict + 0.05 ร— cross

All raw cosine similarities are percentile-scaled (p5โ†’0, p95โ†’1) to normalize distributions across entities. Computed in PostgreSQL via pgvector with HNSW indexes (m=16, ef=64).

Color Science

Every model and backdrop image undergoes pixel-level color analysis in multiple color spaces:

Dominant Color Extraction

OKLab k-means clustering (5 clusters, 10K pixel samples, 32 iterations) extracts the dominant color palette. Results stored as OKLCH (Lightness, Chroma, Hue) and HSV.

Monochrome Detection

A 5-stage algorithm in OKLab (perceptually uniform color space) determines whether an image is monochromatic:

  1. Gray gate โ€” if p95(chroma) < 0.03, classify as near-grayscale
  2. Hue concentration (R) โ€” circular mean of (a, b) vectors; Rโ†’1 = single hue dominates, Rโ†’0 = scattered
  3. Perpendicular spread โ€” p95 deviation from dominant hue axis (threshold: 0.03)
  4. Outlier fraction โ€” max 2% pixels may strongly deviate
  5. Final verdict โ€” monochromatic if all checks pass

Color Separation

For backdrop โ†” model pairs, we measure how distinguishable the model is against its backdrop. Dual-space blend: 70% HSV (hue 50%, saturation 25%, value 25%) + 30% OKLCH with chroma-gated hue distance (low-chroma colors have unstable hue โ†’ gate at C=0.08).

Visual Weight & Rhythm

Rule-based tags from OKLCH values: weight (light/medium/heavy) from 0.5ร—(1-L) + 0.3ร—(C/0.15) + 0.2ร—contrast, and rhythm (calm/dynamic/chaotic) from RMS image contrast thresholds.

Semantic Tagging

Every attribute (model, symbol, backdrop) receives automatic semantic tags derived from LLM-generated keywords + color analysis:

Mood (8 values)
cheerful ยท sad ยท romantic ยท mysterious ยท playful ยท serene ยท intense ยท neutral
Style (8 values)
strict ยท naive ยท gloomy ยท playful ยท minimalist ยท elaborate ยท retro ยท modern
Weight (3 values)
light ยท medium ยท heavy
From OKLCH lightness + chroma
Rhythm (3 values)
calm ยท dynamic ยท chaotic
From RMS image contrast

Mood and style use keyword-matching against 150+ trigger words per category, with LLM category fallback (e.g., "animal" โ†’ playful, "celestial" โ†’ serene + mysterious). Tags power achievements like Mood Curator (80%+ dominant mood) and Chaotic Energy (7+ distinct moods).

3-Layer Search Pipeline

Searching "ๆฐธ้ ใฎ็‚Ž" (Japanese for "eternal flame") triggers a cascading pipeline:

0
Translation
MADLAD-400 (3B params, Google Research, CTranslate2 int8) translates 400+ languages โ†’ English. "ๆฐธ้ ใฎ็‚Ž" โ†’ "eternal flame"
1
Regex Entity Matching
Word-boundary regex across name, slug, theme, artist, LLM keywords, text metadata. Differentiated scoring: name/slug (1.0) โ†’ artist/theme (0.95) โ†’ keywords (0.90) โ†’ text_meta (0.85).
2
Semantic HNSW Search
BGE-M3 (1024d) text embedding โ†’ cosine similarity against models, symbols, collections. HNSW index (m=16, ef=64). Threshold: 0.55.
3
Cross-Modal SigLIP Search
SigLIP So400m/14 (1152d) text โ†’ image space. Finds models and symbols whose images match the query text. Threshold: 0.50.

Entity relevance = max(similarity) across all layers. Gift scoring weights matched attributes: collection +0.03, model +0.04, symbol +0.01, backdrop +0.005. Results cached in Redis (5-min TTL) for instant repeat queries.

AI Model Stack

SigLIP So400m/14
400M params ยท 1152d vectors
Cross-modal textโ†”image search. Layer 3 of the search pipeline and cross-modal compatibility scoring.
BGE-M3 (multilingual-e5-base)
330M params ยท 1024d vectors
Semantic text search. Layer 2 search + semantic compatibility between entities.
MADLAD-400-3B-mt
3B params ยท CTranslate2 int8
Multilingual translation (400+ languages โ†’ English) for search queries.
DINOv2 ViT-L/14
300M params ยท 1024d vectors
Visual feature extraction. Powers visual compatibility scoring between entities.
CLIP ViT-L-14
400M params ยท 768d vectors
Legacy image embeddings for color-aware image analysis and aesthetic scoring.
ALS Collaborative Filtering
Custom ยท 64d vectors
Taste profile embeddings for personalized discovery feed recommendations.

Total: ~4.8B parameters across 6 models. SigLIP, BGE-M3, and MADLAD-400 run in a dedicated semantic-encoder microservice (7GB RAM, 4 CPU). All vector operations use pgvector HNSW indexes in PostgreSQL.

Achievement System

35+ achievements across 8 categories, many with repeatable instances and diminishing-return scaling (log2, sqrt):

Size & Diversity
Whale (500), Leviathan (1000+), World Tour (50 collections)
Rarity
Rare (p95), Epic (p99), Mythic (top 10), repeatable
Monochrome
5-tier: Lover โ†’ Master โ†’ Beast โ†’ Lord โ†’ God. Lord/God can be lost
Contrast
5-tier same as monochrome. Based on color separation score
Serial Numbers
Genesis (#1), Palindrome, Meme Lord, repeatable
Compatibility
Perfect Match (โ‰ฅ0.95), Triple Harmony (all โ‰ฅ0.90)
Color Spectrum
Rainbow (6 hues), Full Spectrum (12), Pastel Dream, Neon Rush
Inscriptions
Love Collection, Funny Bone, Storyteller (200+ chars)

Monochrome/Contrast Lord and God tiers are loss-based โ€” owning a single non-qualifying gift revokes the achievement. This creates meaningful commitment to a collecting strategy.

Infrastructure

153+
API Endpoints
55
Frontend Pages
44
Database Tables
19
Background Workers
55
SQL Migrations
~46K
Lines of Code

Tech Stack

Next.js 15React 19FastAPIPython 3.11PostgreSQL 15pgvectorRedis 7ClickHouseDockerMinIOTON BlockchainTelegram MTProtoSigLIPBGE-M3DINOv2MADLAD-400CTranslate2OKLab/OKLCHHNSWTailwind CSS
โ–’โ–ˆโ–ˆโ–ˆ โ–’โ–ˆโ–ˆโ–ˆ โ–’โ–ˆโ–ˆโ–ˆ โ–’โ–ˆโ–ˆโ–ˆ    โ–’โ–ˆโ–ˆโ–ˆ โ–’โ–ˆโ–ˆโ–ˆ    โ–’โ–ˆโ–ˆโ–ˆ
โ–’โ–ˆโ–ˆโ–ˆ โ–’โ–ˆโ–ˆโ–ˆ โ–’โ–ˆโ–ˆโ–ˆ โ–’โ–ˆโ–ˆโ–ˆ    โ–’โ–ˆโ–ˆโ–ˆ โ–’โ–ˆโ–ˆโ–ˆโ–ˆ  โ–’โ–ˆโ–ˆโ–ˆโ–ˆ
โ–’โ–ˆโ–ˆโ–ˆ โ–’โ–ˆโ–ˆโ–ˆ โ–’โ–ˆโ–ˆโ–ˆ โ–’โ–ˆโ–ˆโ–ˆ    โ–’โ–ˆโ–ˆโ–ˆ โ–’โ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–’โ–ˆโ–ˆโ–ˆโ–ˆโ–ˆ
โ–’โ–ˆโ–ˆโ–ˆ โ–’โ–ˆโ–ˆโ–ˆ โ–’โ–ˆโ–ˆโ–ˆ โ–’โ–ˆโ–ˆโ–ˆโ–ˆ  โ–’โ–ˆโ–ˆโ–ˆโ–ˆ โ–’โ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆ
โ–’โ–ˆโ–ˆโ–ˆ โ–’โ–ˆโ–ˆโ–ˆ โ–’โ–ˆโ–ˆโ–ˆ โ–’โ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆ โ–’โ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆ
โ–’โ–ˆโ–ˆโ–ˆ โ–’โ–ˆโ–ˆโ–ˆ โ–’โ–ˆโ–ˆโ–ˆ  โ–’โ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆ โ–’โ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆ
โ–’โ–ˆโ–ˆโ–ˆ โ–’โ–ˆโ–ˆโ–ˆ โ–’โ–ˆโ–ˆโ–ˆ   โ–’โ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆ โ–’โ–ˆโ–ˆโ–ˆโ–’โ–ˆโ–ˆโ–ˆโ–’โ–ˆโ–ˆโ–ˆ
โ–’โ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆ         โ–’โ–ˆโ–ˆโ–ˆ โ–’โ–ˆโ–ˆโ–ˆ โ–’โ–ˆ โ–’โ–ˆโ–ˆโ–ˆ
โ–’โ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆ โ–’โ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆ โ–’โ–ˆโ–ˆโ–ˆ    โ–’โ–ˆโ–ˆโ–ˆ
โ–’โ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆ โ–’โ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆ โ–’โ–ˆโ–ˆโ–ˆ    โ–’โ–ˆโ–ˆโ–ˆ
โ–’โ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆ โ–’โ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆโ–ˆ  โ–’โ–ˆโ–ˆโ–ˆ    โ–’โ–ˆโ–ˆโ–ˆ

Ein Produkt des Kollektivs Schum