How I Built an Oreo Generator with 1.1 Sextillion Combinations

TLDR

I built a web app that generates 1.1 sextillion possible Oreo combinations from 406,718 flavors. Try the generator or watch the video where I explain the process. Everything from Chocolate Chip Oreo to Pokemon Sriracha Danny DeVito Oreo.

---

Project Summary

Metric	Value
Total combinations	1.1 sextillion
Flavors in database	406,718
Raw records processed	31 million
Manual review time	3-4 days

The math is straightforward combinatorics. The real challenge was building a comprehensive flavor database.

---

Table of Contents

• The Mathematical Foundation
• Defining Flavor Categories
• Building the Flavor Database
  • Phase 1: Data Acquisition
  • Phase 2: Automated Processing & Filtering
  • Phase 3: Manual Review Interface
• The Combination Logic
• Building the Generator
• The Scoring System
• Using the Generator
• The Numbers
• Try It Yourself

---

The Mathematical Foundation

The formula for counting combinations is the binomial coefficient:

function combination(n: number, k: number): bigint {
  if (k > n) return 0n;
  if (k === 0 || k === n) return 1n;

  let result = 1n;
  for (let i = 1; i <= k; i++) {
    result = (result * BigInt(n - i + 1)) / BigInt(i);
  }
  return result;
}

This calculates C(n,k): the number of ways to choose k items from n total items without replacement, where order doesn't matter and no item appears twice.

For Oreos with 1-4 flavors from n total flavors:

C(n,1) + C(n,2) + C(n,3) + C(n,4)

The implementation uses BigInt because JavaScript numbers lose precision beyond 2^53. Standard numbers can't accurately represent values like 1.1 sextillion.

First, I needed to find n: the total number of possible flavor components.

---

Defining Flavor Categories

Looking at real Oreo releases, I noticed they fall into three distinct categories:

1. Food Flavors

Traditional flavors that you can actually taste: Birthday Cake, Red Velvet, Peanut Butter, Mint.

2. Food Brand Collaborations

Branded food products that bring their own flavor profiles: Reese's Peanut Butter Cup, Chips Ahoy!, Sour Patch Kids, Swedish Fish.

3. Non-Food Brand Collaborations

These are where Oreo gets creative with pure branding: Supreme, Game of Thrones, Star Wars, Lady Gaga, Selena Gomez.

When Oreo releases a Pokemon Oreo, they're not claiming it tastes like Pikachu (hopefully). It's about the brand association.

---

Building the Flavor Database

I built the database in three phases: raw data acquisition, automated processing and filtering, and manual review.

---

Phase 1: Data Acquisition

No single dataset contains every possible flavor. I searched for comprehensive food databases, but most had significant gaps or poor organization.

Wikipedia's food lists turned out to be surprisingly useful: easy to scrape and full of unique, high-quality entries.

Eventually, I assembled data from food databases, brand repositories, entertainment franchises, and celebrity databases.

Food Flavors

I gathered raw source files containing 427,344 total records from multiple food databases:

Source	Count	Description
Wikipedia Food Lists	261,547	Raw file entries from comprehensive food lists
USDA Foundation Foods Database	155,244	Raw database rows (nutritional data, metadata, duplicates)
Food Identicon	4,999	Community-maintained ingredients
Wikidata	3,222	Dishes, variations, and food variants via SPARQL queries
CORGIS Ingredients Dataset	2,332	Academic dataset entries
Total Raw Food Records	427,344

Food Brands

Source	Count	Description
Open Food Facts	4,000,000	1.1GB CSV database with products
Wikidata Food Brands	1,101	Structured food and beverage brands via SPARQL queries
Total	4,001,101

Non-Food Brands

I broke non-food brands into types to track totals for each category.

Entertainment Characters & Franchises

Source	Count	Description
Marvel Characters	16,376	Characters from FiveThirtyEight dataset
DC Characters	6,896	Characters from FiveThirtyEight dataset
Disney Characters	9,820+	From Disney API plus manual curation
Pokemon	900+	Pokemon with complete stats
Anime	525+	Dragon Ball (209), One Piece (316+)
Gaming	238+	Wikipedia gaming/Nintendo characters
Franchises	563	Wikipedia (425) + Wikidata (138) franchises
Total	35,318

Sports Teams

Source	Count	Description
Wikipedia Sports	1,190	Professional teams
Wikidata Sports	1,500	Sports entities with metadata
Total	2,690

Music & Celebrities

Source	Count	Description
MTV Artists	300+	Artists with social media and genre data
IMDb Names	14,681,812	Names for celebrity extraction
IMDb Titles	11,877,454	Titles
Manual Celebrities	99	Manually curated
Total	26,559,665

Non-Food Brands Summary

Category	Count
Entertainment Characters & Franchises	35,318
Sports Teams	2,690
Music & Celebrities	26,559,665
Total Non-Food Brands	26,597,673

Total Raw Sources

Category	Count	Description
Food Flavors	427,344	Raw source records
Food Brands	4,001,101	Raw product records
Non-Food Brands	26,597,673	Raw source records
Total Raw Records	31,026,118

---

Phase 2: Automated Processing & Filtering

The 31 million raw source records needed extraction and processing. First, I extracted actual terms from the raw files (pulling food names from nutritional databases, brand names from product catalogs, character names from entertainment datasets). Then I filtered out technical codes, duplicates, and non-food terms. Each category required different processing strategies.

Food Flavors

Starting from 427,344 raw source records, I extracted food-related terms, then filtered to remove technical codes, duplicates, and non-food terms:

For example, USDA entries like Beef, eye of round roast, raw (ER37-R-23) were filtered out as too technical.

Filtering out technical codes was straightforward. But duplicates were harder: what even counts as a duplicate?

The Duplicate Problem

Simple flavors like potato and potatoes are obviously duplicates. But what about potato vs. baked potato vs. mashed potato vs. potato salad?

Without clear rules, I might accidentally filter out valid entries that should have been kept.

If Oreo released both Baked Potato Oreo and Mashed Potato Oreo, I think you'd expect them to taste completely different (well... maybe not completely but they'd have a different vibe.) Therefore, the way a food is prepared can change whether it's a duplicate or a unique flavor.

Preparation methods aren't the only thing that creates distinct flavors. Complete dishes combine multiple ingredients into something entirely new. Pizza Pringles exist as a product, which means pizza is its own distinct flavor, not just cheese or tomato or bread. A Pizza Oreo would taste different from a Cheese Bread Oreo.

Three-Tier Classification System

The solution was a three-tier classification system:

Tier	Description	Examples
Base	Core ingredients	potato, apple, chocolate
Variation	Prepared forms	mashed potato, apple pie, dark chocolate
Dish	Complete dishes	potato salad, apple strudel, chocolate cake

With these rules defined, I could now consistently decide what to keep and what to filter. Potato and potatoes are duplicates (keep one). But potato, baked potato, and potato salad are all unique entries across different tiers (keep all three).

Now that we have the correct classification system in place let's see the complete filtering process:

#	Step	Before	After	Description
1	Extraction	427,344	309,716	Extracted food terms from raw source files (e.g., pulled food names from 155k USDA nutritional data rows)
2	Bad Entry Filtering	309,716	229,577	Removed technical codes, overly long/short terms, entries with numbers
3	Non-Food Filtering	229,577	228,847	Eliminated non-food terms using keyword patterns
4	Classification	228,847	228,847	Sorted into base/variation/dish categories
5	Deduplication	228,847	66,745	Removed duplicates and normalized variations
Result		427,344	66,745

Food Brands

The 4 million products in Open Food Facts don't represent 4 million unique brands. Many products share the same brand, just in different sizes or formats.

For example, Open Food Facts contains thousands of Coca-Cola products:

• Coca-Cola Classic 330ml can
• Coca-Cola Classic 2L bottle
• Coca-Cola Classic 500ml bottle
• Coca-Cola Zero 330ml can
• Coca-Cola Zero 2L bottle

These are all different products, but they don't all represent different flavors. I needed to extract brand variants that actually matter for taste. Coca-Cola Classic and Coca-Cola Zero taste different, so I kept them as separate flavors. But the size (330ml vs 2L) doesn't change the flavor, so I only kept one. Extracting unique brands from 4M products gave me 307,674 brand variants.

Here's the complete filtering process:

#	Step	Before	After	Description
1	Extraction	4,001,101	307,674	Extracted unique brand names from product records, deduplicated across products
2	Filtering & Cleaning	307,674	285,441	Removed 22,233 entries: pure punctuation, technical codes, URLs, entries with no letters
3	Deduplication	285,441	282,347	Merged 3,094 spelling variants like "Dr. Pepper" → "Dr Pepper", "Ben and Jerry's" → "Ben & Jerry's"
4	Database Import	282,347	281,996	Final quality filters during import
Result		4,001,101	281,996

Non-Food Brands

The real challenge with non-food brands: what even counts as a flavor? Is Microsoft Excel a flavor? Is Ikea a flavor?

I focused on recognizability. We don't want entries so obscure that nobody gets the reference. A Stormtrooper who bonked his head in Star Wars A New Hope Oreo might be fun, but it's too specific.

A Lady Gaga Oreo or Pokémon Oreo makes sense because people actually know them.

This constraint meant aggressive filtering across 26 million raw entries:

#	Step	Before	After	Description
1	Extraction & deduplication	26,597,673	150,404	Extracted individual names and deduplicated to unique terms like "Batman", "The Avengers", "Robert Downey Jr."
2	Cleaning & classification	150,404	149,884	Removed invalid entries
3	Deduplication	149,884	131,155	Removed more duplicates
4	Normalization	131,155	130,712	Standardized formats
5	Database import	130,712	130,989	Final import to database
Result		26,597,673	130,989

Total After Processing

Category	Before	After	Reduction
Food Flavors	427,344	66,745	84.4%
Food Brands	4,001,101	281,996	93.0%
Non-Food Brands	26,597,673	130,989	99.5%
Total	31,026,118	479,730	98.5%

---

Phase 3: Manual Review Interface

After all the automated processing, I still had 479,730 entries that needed manual curation. I built a comprehensive review interface directly into the generator to go through each entry one by one.

The manual review interface: approve, reject, or flag for cleanup

Over 3-4 days, I reviewed every entry. Bad entries like Broccoli, raw, OCEAN MIST, 4 bunches; Product of USA (NC1) - NFY0905IX got rejected as too technical. Legitimate flavors like Danny DeVito got approved. Entries with weird formatting got flagged for cleanup.

The combination of automated filtering and manual review reduced the database from 479,730 entries to 404,539 flavors, before adding the dictionary.

Final Flavor Count

Manual review filtered each category:

Category	Before	After	Rejected
Food Flavors	66,745	33,856	32,889
Food Brands	281,996	251,179	30,817
Non-Food Brands	130,989	119,504	11,485
Total	479,730	404,539	75,191

The Dictionary Addition

After reviewing 66,745 food entries, I had 33,856 approved food flavors. But I couldn't shake the feeling I was missing potential flavors.

Somewhere out there was the platonic ideal of an Oreo flavor and I didn't want to miss it.

So I did the only rational thing: I read the dictionary. I processed every entry, filtered for food-related words, and extracted 2,179 additional food terms. Adding these to the existing 33,856 gave me 36,035 final food flavors.

Category	Count	Description
Food Flavors (After Manual Review)	33,856
Dictionary Addition	+2,179	Food-related words from dictionary processing
Final Food Flavors	36,035

Food Flavors breakdown by tier:

Tier	Before	After	Rejected
Base	39,114	24,275	14,839
Variation	9,803	7,449	2,354
Dish	17,828	4,311	13,517
Total Food Flavors	66,745	36,035	30,710

Final Total Across All Categories

Category	Count
Food Flavors	36,035
Food Brands	251,179
Non-Food Brands	119,504
Total Final Flavors	406,718

---

The Combination Logic

With 406,718 total flavor components, I needed to set a reasonable limit on combinations.

The problem: you can't just stack flavors endlessly. Mix too many paint colors together and you get brown. Mix too many flavors and you get the same result: an indistinguishable mess.

A Bacon Mango Doritos Soy Sauce Mint Tom Hanks Pickle Oreo isn't innovative, it's just noise.

Your taste buds can't process that many competing signals. So I capped it at 4 flavors per Oreo max.

The calculation counts all possible Oreos with 1, 2, 3, OR 4 flavors (because an Oreo needs at least one flavor):

k Value	Combinations	Description
k=1	406,718	Just one flavor
k=2	82,709,562,403	Two flavors combined
k=3	11,213,100,794,099,516	Three flavors combined
k=4	1,140,134,072,368,046,162,485	Four flavors combined
Total	1,140,145,285,551,550,231,122

Total possible Oreo combinations: 1,140,145,285,551,550,231,122

The 4-flavor combinations (k=4) make up 99.9990165% of all possibilities. Each additional slot increases the combinations factorially, which is why the jump from k=3 to k=4 is so massive.

Exponential growth: k=4 combinations dwarf everything else

---

Building the Generator

Now that I've mapped 1.1 sextillion Oreo combinations, I need to build a generator that lets people actually explore them.

The obvious approach doesn't work: you can't pre-generate and store all combinations. My computer has 1 terabyte of storage. Even at just 1 byte per combination, the total storage needed is:

1,140,145 petabytes
Storage needed at 1 byte per combination

So... obviously we can't do that.

Instead, the generator picks random flavors from the database on demand, calculating combinations in real-time.

The first version was straightforward: SELECT 4 random flavors, display them. Done.

Except it was boring. Without scores, every combination felt equally meaningless.

---

The Scoring System

I added a scoring system with three components: edibility, harmony, and novelty. Each combination gets rated 0-100% based on how the flavors work together.

Problem solved, right?

Wrong. Most combinations still clustered around mediocre. Strawberry Vanilla Cookie Cream Oreo scores 52%. Chocolate Banana Caramel Mint Oreo scores 48%. The mathematical average of random ingredients is... average.

Nobody wants to browse through endless 45-55% scores. The generator needed something more aggressive.

Flavor Tiers

The scoring system works by classifying every flavor in the database into one of four tiers:

Tier	Description	Examples
God-Tier	Classic dessert ingredients	vanilla, chocolate, cookie, peanut butter, caramel
Great	Popular fruits and spices	strawberry, mint, cinnamon, coffee
Weird	Savory foods that could work	bacon, pizza, cheese, popcorn
Terrible	Things that should never be Oreos	soap, garbage, gasoline

Each tier affects how the three components get calculated:

Edibility

Tier	Range	Notes
God-tier	85-100	Classic dessert flavors
Great	75-95	Popular fruits and spices
Weird	80-100 or 5-35	Unpredictable (20% chance high, 80% chance low)
Terrible	1-15	Should never be Oreos

Harmony

Pairing Type	Range	Notes
Food-food	40-100	Natural flavor combinations
Food-brand	30-90	Mixed compatibility
Brand-brand	1-80	Wildcards, unpredictable

Novelty

Category	Range	Notes
Non-food brands	50-100	Highest novelty
Food brands	30-90	Moderate novelty
God-tier foods	5-70	Lowest novelty (too common)

A combination of four god-tier flavors might score high on edibility and harmony but low on novelty. Mix in terrible flavors and you get the opposite: high novelty, terrible edibility.

But most combinations still land around 50%. The breakthrough was an anti-boring algorithm that pushes boring middle scores toward the extremes.

Anti-Boring Algorithm

When combinations score 45-55%, there's a 40% chance the algorithm pushes them to extremes:

private avoidBoringScores(edibility, harmony, novelty, finalScore, randomSeed) {
  if (finalScore >= 45 && finalScore <= 55) {
    // 60% chance: keep boring score
    // 40% chance: make it exciting
    if (this.seededRandom(randomSeed + 2000) < 0.6) {
      return { edibility, harmony, novelty, finalScore };
    }

    // Push to extreme: 60-90% if above 50, or 10-40% if below
    const targetScore = finalScore > 50
      ? 60 + this.seededRandom(randomSeed + 2001) * 30
      : 10 + this.seededRandom(randomSeed + 2002) * 30;

    // Adjust all three components to hit target
    const adjustment = targetScore - finalScore;
    const adjustmentPerComponent = adjustment / 3;

    return {
      edibility: edibility + adjustmentPerComponent,
      harmony: harmony + adjustmentPerComponent,
      novelty: novelty + adjustmentPerComponent,
      finalScore: (edibility + harmony + novelty) / 3
    };
  }

  return { edibility, harmony, novelty, finalScore };
}

This creates more polarizing results: god-tier (80%+) or terrible (20%-), with fewer "meh" scores.

---

Using the Generator

The Interface

The UI gives you four empty slots. You drag flavor types (Food, Brand Food, or Non-Food Brand) into each slot. You can fill all four slots, or just one. Click "Generate" and the backend randomly picks flavors matching your constraints from the database, scores them using the tier-based system and anti-boring algorithm, then returns a combination with scores.

Each generated combination shows:

• Flavor Names: The selected flavors and their types
• Component Scores: Edibility, Harmony, and Novelty (0-100%)
• Overall Score: Final rating from 0-100%
• Analysis Label: "GOD-TIER", "CHAOTIC MESS", etc.

You can copy the result as an image or share it to Twitter/Bluesky.

Drag flavor types into slots, generate, and see the scored results

Achievements

The generator tracks your session and unlocks achievements:

Milestone:

• First Bite: Your first generation
• Flavor Explorer: Generate 10 combinations
• Combination Master: Generate 100 combinations

Discovery:

• Reality Check: Match an actual Oreo product
• Unique Explorer: Discover 100 unique combinations

Special:

• Perfect Harmony: Score 90% or higher
• Brave Soul: Score less than 10%
• Speed Demon: Generate 10+ per minute

Session stats show how many you've generated, your best score, generation rate per minute, and how long you've been generating.

Tech Stack

Backend:

• NestJS
• TypeScript
• SQLite
• TypeORM for queries

Frontend:

• Vite
• React
• TypeScript
• Lucide React for icons
• Canvas API for image generation

---

The Numbers

Final stats from the actual implementation:

• 406,718 total flavors in database
• 1,140,145,285,551,550,231,122 total possible combinations (all 1-4 flavor Oreos)

From 31 million raw records to 1.1 sextillion combinations, the Oreo Generator maps the entire flavor possibility space. Somewhere out there is the Platonic Ideal of an Oreo Flavor, and now you can find it.

---

Try It Yourself

Try the Oreo Generator yourself or watch the video walkthrough to see the complete story.