Fable — Allergen-Aware Recipe Discovery

Safe ingredients. Bold flavours. Food for everyone.

Fable is an allergen-aware recipe discovery app powered by Epicure — the largest multilingual food embedding model ever built (4.1M recipes, 7 languages, 1,790 ingredients). It uses ingredient embeddings to find flavour-matched pairings and generates novel, restaurant-quality recipes tailored to your dietary restrictions.

Built for the H0 Hackathon (AWS + Vercel, May–June 2026).

Tech Stack

Layer	Technology
Frontend	Next.js 16, React 19, Tailwind CSS v4, Framer Motion
Deployment	Vercel
Database	AWS DynamoDB
Embeddings	Epicure Core (1,790 ingredients × 300 dimensions, cosine similarity)
Recipe generation	Anthropic Claude (`claude-sonnet-4-6`) with prompt caching; `claude-haiku-4-5` for recipe brief
Allergen data	EU Big 14 truth table — 1,790 ingredient classifications, O(1) lookup
Package manager	pnpm
Lambda	AWS Lambda (`nodejs24.x`) — DynamoDB Streams feedback processor · ingredient insights writer · Claude Vision ingredient scanner
Testing	Jest 29, ts-jest, React Testing Library — 575 tests across 28 suites; 12 Lambda tests (node:test)

Local Development

# Install dependencies
pnpm install

# Copy environment variables
cp .env.example .env.local  # then fill in the values below

# Create DynamoDB tables (one-time setup)
pnpm setup:dynamodb

# Seed the ingredient insights table with realistic starter data
pnpm seed:insights

# Start the dev server
pnpm dev

Environment Variables

Variable	Description
`AWS_REGION`	AWS region (e.g. `eu-west-2`)
`AWS_ACCESS_KEY_ID`	AWS access key
`AWS_SECRET_ACCESS_KEY`	AWS secret key
`ANTHROPIC_API_KEY`	Anthropic API key for recipe generation
`VISION_LAMBDA_URL`	API Gateway URL for the `fable-vision-ingredient-scanner` Lambda

Note: the project uses pnpm. Running npm install will create a package-lock.json that conflicts with pnpm-lock.yaml and break the Vercel build.

Shipped Features

Allergen System

EU Big 14 allergen picker with emoji cards
Custom allergen search across all 1,790 Epicure ingredients
Allergen codes mapped to EU standard (milk not dairy, tree_nuts not tree-nuts)
Settings accessible at any time via the header cog; profile persisted in DynamoDB
Quick-add ingredient chips automatically exclude allergen-flagged items and backfill from a wider pool

Ingredient Input

Searchable across all 1,790 Epicure ingredients with autocomplete
Staging panel on ingredient selection — set subtype, quantity, unit, and date before confirming
- Subtype — e.g. "breast", "ribeye", "baby"; appends to display name and improves recipe accuracy
- Quantity + unit — pieces, grams, kg, ml, litres, tbsp, tsp, cups
- Date mode — "Use by date" (user-entered) or "Bought date" (expected expiry auto-calculated from shelf-life table)
Kitchen area badges on every ingredient tag — 🧊 Fridge, ❄️ Freezer, 🗄️ Cupboard, 🏠 Pantry
Expiry warnings on ingredient tags — amber at 2 days, red with "Use today!" at 0–1 days
Quick-add chips for popular ingredients (adapts to Safe Foods list when mode is active)
Meal type filter — Snack, Starter, Main Course, Dessert
Cook time filter — Quick (<30 min), Medium (30–60 min), Slow Cook (60 min+)
Cuisine filter — 13 cuisines + "Surprise me", horizontal-scrollable chip row, single-select with "Any cuisine" default
Occasion filter — Weeknight, Dinner Party, Street Food, Comfort Food, Packed Lunch, Romantic Dinner, Meal Prep, Celebration
Servings stepper — +/− control for 1–12 people (default 2); quantities in the prompt are scaled accordingly
Kitchen equipment — collapsible "What equipment do you have?" section with checkboxes for Hob, Oven, Microwave, Air Fryer, Slow Cooker, Pizza Oven, Barbecue/Grill, Instant Pot; Hob and Oven on by default; preference persisted permanently to DynamoDB
Use my kitchen only toggle — constrains recipe generation to exactly the ingredients added; skips Epicure pairings and adds a hard prompt constraint so Claude adds nothing extra
Ingredient list persisted in DynamoDB with debounced auto-save; old string-array profiles migrated automatically

Photo Ingredient Recognition

Camera icon alongside the kitchen search bar — taps open the device camera or photo library
Photo is base64-encoded and sent to POST /api/scan-ingredients (Next.js proxy → fable-vision-ingredient-scanner Lambda)
Lambda calls Claude Vision (Haiku 4.5) with a structured prompt to identify ingredients and infer storage area from visual cues (fridge, freezer, cupboard, pantry)
Claude's ingredient names are fuzzy-matched against all 1,790 Epicure keys: exact match, prefix match, then token-overlap scoring; anything below threshold is excluded
confident: false flagged when Claude expressed uncertainty or when the fuzzy match score is below the 0.8 high-confidence threshold
Full review screen before anything lands in the kitchen: inferred storage area shown as a badge with one-tap editing; changing area applies to all ingredients at once
Each ingredient row shows the display name, Epicure key beneath, an Uncertain badge when confidence is low, and an In kitchen badge when already present — pre-deselected to prevent duplicates
Confirmed ingredients written via setIngredients (triggers the existing debounced DynamoDB auto-save — no new endpoint)
Done with nothing selected = Cancel; no write made
Error toasts for Lambda timeout, no ingredients found, and network failure (Sonner, now mounted in root layout)

Recipe Generation

Show Pairings — Epicure similarity search surfaces safe, flavour-matched ingredients
Generate Recipe — Two-step agentic flow: Claude Haiku reasons about taste history and writes a RecipeBrief (step 1), then Claude Sonnet generates the recipe guided by the brief (step 2)
Recipe Brief Card — Shown during recipe generation: displays the dish direction, reasoning, novelty note, and rotating cooking hints. For guests or users with insufficient history, shows the default Fable loading hints
Ingredients sorted by expiry date (ascending) before being passed to Claude — expiring items get used first
Claude receives rich descriptions (e.g. "2 pieces Chicken Breast (Epicure: chicken)") for cut-accurate recipes
Ingredient quantities displayed rounded — whole-unit items (pieces, cloves, fillets) always shown as integers
Recipes respect allergen profile, meal type, and cook time
Prompt caching on the system prompt (~90% cost reduction on repeated calls)
Food-themed gradient hero on every recipe card with title-hash colour variation across five palettes

Drink Pairings

Automatically suggested after every recipe is generated
Top 3 recipe ingredients by quantity run through Epicure cosine similarity to find the closest-matching beverages
Filtered against a curated list of 55 drinking beverages (cooking wines, vinegars, and non-drink items excluded)
Allergen profile applied — e.g. milk-allergic users will not see milk or oat milk suggestions
Context-aware emoji per drink type: 🍵 tea · ☕ coffee · 🥛 milk · 🍺 beer/cider · 🍷 wine · 🧃 juice · 🍸 spirits

Recipe Feedback

👍 / 👎 buttons on every generated recipe — feedback saved immediately to DynamoDB, then an optional survey panel appears
Feedback survey — four-section panel of tappable multi-select chips:
- ✨ Highlight of the dish — select ingredients you loved (chips from the generated recipe)
- 🚫 Would leave out — select ingredients you'd skip (same list; mutual exclusion with Section 1)
- 👌 What worked — Perfect complexity · Great cuisine choice · Right amount of ingredients · Quick to make
- 😬 What didn't — Too complex · Too simple · Wrong cuisine vibe · Too many ingredients · Took too long
- Skip dismisses without any API call; Done with nothing selected is treated as Skip
Feedback stored in DynamoDB (fable-feedback table): userId, recipeId, liked, reasons, notes, recipeTitle, recipeIngredients, allergenProfile, timestamp, surveyResponse (optional)
Survey responses persisted via PATCH /api/feedback — updates the existing record, never overwrites base feedback
Recent disliked patterns and ingredients loaded at session start and injected into the Claude prompt — future recipes actively avoid them
Survey-informed generation — /api/generate-recipe reads survey signals from the last 20 feedback records and injects them into the Claude prompt (threshold-gated at 3+ records):
- ingredientsHighlighted boosts ingredient preference score by 1.5×; ingredientsSkipped reduces by 1.5× (additive on top of base like/dislike scores)
- recipePositives and recipeNegatives aggregated by buildPreferenceProfile; signal strings appearing 2+ times converted to prompt clauses via formatSignalsToClauses → RECIPE_SIGNAL_MAP and injected as soft guidance (e.g. "Keep the method simple", "Keep cook time short")
- Format signals also surfaced in the taste profile card as "Your preferences" chips with neutral display labels (e.g. 'Too complex' → 'simpler recipes')
Real-time preference signals — DynamoDB Stream on fable-feedback triggers fable-feedback-stream-processor Lambda on every write; one preferenceSignals entry per ingredient is appended to fable-users automatically (event-driven, no polling)

Safe Foods Mode

For users with MCAS, severe allergies, or highly restricted therapeutic diets.

User builds an explicit list of every ingredient they can safely eat
Recipe generation is strictly constrained to that list — no unlisted ingredients, substitutions, or garnishes
Post-generation validation strips any ingredient Claude included outside the safe list
If no liquid is in the safe list, Claude uses the placeholder "liquid of choice" with an in-step note; same for salt/seasoning
Quick-add chips on the ingredient screen switch to the user's safe foods list
Mode indicator badge in the app header; toggle in allergen settings
Safe ingredients and mode preference persisted in DynamoDB

Dark Mode

🌙 / ☀️ toggle in the header alongside the settings cog
Toggle also available in Allergen Settings
Preference persisted to DynamoDB — carries across devices
Full dark: variant support via Tailwind and the existing .dark CSS variable theme

Nutritional Information (Macros)

"Show nutritional information" toggle in Allergen Settings — off by default
Note displayed beneath the toggle: "Calorie and macro information is hidden by default out of respect for users in eating disorder recovery."
When on, Claude estimates calories, protein, carbs, and fat per serving as part of recipe generation
If the toggle is turned on after a recipe has already been generated, macros are fetched on demand via /api/macros (Claude Haiku) and patched into the displayed recipe without regenerating it
Displayed under the label "Estimated nutritional information" as a four-cell row (Calories · Protein · Carbs · Fat) with a disclaimer: "Estimates based on ingredients and quantities — consult a nutritionist for precise values."
Toggle preference persisted in DynamoDB on the user profile

Collections

Saved recipes organised into named collections, persisted in DynamoDB
"All Saved" and "Collections" tabs on the saved screen
Collection cards show a gradient colour-strip mosaic preview of the first 3 recipes inside
Bookmark icon on every saved recipe card opens an add-to-collection modal — toggle recipes in/out of any collection with a single tap
"New collection" creation inline in both the modal and the collections tab
Collection detail view shows the full recipe grid; removing a recipe removes it from the collection only (not from saved)
Collections load on session start alongside profile and saved recipes

Substitutes Mode

Allergen-safe ingredient substitution using Epicure embeddings, with full recipe adaptation.

From my kitchen — tap any kitchen ingredient to see the top 3 safe substitutes, scored by a weighted combination of similarity to the original (60%) and cosine fit to the rest of the dish (40%)
From a recipe — paste a full recipe or enter an ingredient list; Claude extracts the ingredients, then each one is automatically checked:
- ✅ In kitchen — ingredient is available, included as-is
- 🔄 Allergen — contains a user allergen; best safe kitchen substitute auto-suggested via embedding search
- ⚠️ Missing — not in kitchen, not an allergen; included in the adapted recipe anyway
Substitution plan displayed as a formatted list before committing, with quantities where known (e.g. 🔄 Butter (2 tbsp) → Olive Oil (82% match) from your kitchen)
Cook with these substitutions — builds the adapted ingredient list and generates a full recipe via Claude, maintaining the spirit of the original dish
Functional category scoring — same-category substitutes get a +0.1 score bonus, different-category a −0.3 penalty; grain ingredients (pasta, rice, flour) are hard-filtered from fat/dairy/liquid targets regardless of embedding score
Expiry-aware ranking (From my kitchen) — kitchen substitutes expiring within 2 days are boosted up the ranking; expiry badge shown on result cards (red for today, amber for 2 days). Boost only applies if the base score is ≥ 45%
Swap icon (↔) on every ingredient row in the generated recipe screen opens Substitutes pre-loaded with that ingredient and the rest of the recipe as context
Find Substitutes button on the ingredients screen; dedicated Substitutes tab in the bottom navigation

Diet & Lifestyle Presets

One-tap diet restriction setup above the EU Big 14 allergen grid.

Four presets: 🌱 Vegan, 🥗 Vegetarian, 🥑 Keto, 🟢 Low-FODMAP — each maps to a curated list of Epicure ingredient keys excluded from recipe generation and pairings
Lactose Intolerance toggle with two sub-modes (expand when enabled):
- Include dairy with reminders — dairy stays in recipes; Claude adds a Lactaid note to the description; a 🥛 banner appears on the recipe screen; dairy kitchen ingredients show a 🥛 indicator on their tag
- Exclude dairy entirely — treats dairy exactly like a milk allergen, filtered from all results and recipe generation
lactoseMode: 'include' | 'exclude' persisted in DynamoDB alongside the toggle flag
Presets stack with EU Big 14 allergens and custom allergen selections; exclusions are computed at call time (effectiveCustomAllergens) without mutating stored preferences
Collapsible section auto-expands on load when any option is active (watches isLoadingProfile to handle async DynamoDB load)
Header subtitle reflects the full restriction picture: "Vegan + 3 allergens active"
Substitutes "From a recipe" allergen mode: ingredients below the 45% combined score threshold show ❌ [ingredient] — contains [allergen], no suitable substitute found — will be omitted rather than the 🔄 swap format

Onboarding Tutorial

A 5-slide introductory slideshow that appears on first launch and is re-launchable from settings.

Shown automatically on first app load by checking fable-onboarding-complete in localStorage; never shown again once dismissed
Full-screen overlay with a dark backdrop dimming the app behind it; slides animate left/right via Framer Motion
Slide 1 — Welcome: Fable logo, food-gradient hero, brand positioning copy
Slide 2 — Allergens: EU Big 14 allergen picker illustration with active state previews
Slide 3 — Your Kitchen: Fridge and Cupboard ingredient card mockup
Slide 4 — Recipe Generation: Generated recipe card with Claude attribution and gradient hero
Slide 5 — Safe Foods Mode: Safe Foods list with green shield and ingredient checklist
Skip button (top-right) dismisses at any point; "Let's go" CTA on the final slide — both set fable-onboarding-complete: true
Tapping anywhere on the slide body advances to the next slide (mobile-friendly)
Dot indicator (pill-shaped active dot) shows current position across all 5 slides
Restart tutorial option in Allergen Settings — clears the localStorage flag and re-shows the overlay
8 unit tests covering: show-on-first-load, suppress-when-seen, skip flag, Let's go flag, restart flag, slide count, dot index range

Navigation & History

Six-tab navigation — Kitchen, Recipe, Discover, Substitutes, History, Saved
Responsive layout — bottom tab bar on mobile (< 768 px); fixed left sidebar (220 px) on desktop (≥ 768 px) with Fable logo/wordmark at the top and stacked icon + label items; main content area shifts right on desktop; same green active-state highlight and bg-card/95 theming at both breakpoints
Recipe tab persists the most recent recipe across navigation
History tab — all recipes generated this session, newest first
Saved tab — hearted recipes persisted in DynamoDB; deletable
Tab visibility — Navigation settings let users hide individual tabs; at least 2 must remain visible; persisted to DynamoDB

Discover Tab

Trending ingredient insights as a dedicated tab (Compass icon), between Recipe and Substitutes.

Your taste profile — personalised card rendered once the user has 5+ feedback signals, showing: ingredients they love, ingredients they avoid, a flavour territory (2–4 ingredients from the intersection of their liked ingredients' embedding-space neighbours), and a "Your preferences" row of format signals that appear 2+ times (e.g. "simpler recipes", "shorter cook time"). Computed from buildPreferenceProfile + deriveFlavourTerritory
Trending for you — top 3 recipe types (cuisine + occasion) trending for the user's allergen profile this week; tapping pre-fills the cuisine and occasion filters and navigates to the ingredient screen. When a user has ≥ 5 signals, each chip also injects their top 3 personally loved ingredient keys as a soft hint into recipe generation
Trending globally — top 5 most-liked ingredients across all users this week
Most loved ingredients — all-time top 6 ingredients for the user's allergen profile, shown with a visual score bar
Trending pairings — top 3 drink + cuisine pairings this week for the user's allergen profile
Section and each sub-section individually toggleable in settings; persisted to DynamoDB
Powered by fable-ingredient-insights — a fifth DynamoDB table aggregated by the Lambda on every liked feedback event
API route /api/insights cached for 1 hour (Next.js route revalidation)
Subtitle shows "Trending this week · global" when the user has no restrictions; "Trending this week · safe for you" with an "Excluding [allergen list]" secondary line when any EU 14 or custom allergens are set — driven by the user's actual restrictions, not by whether their profile matched a seeded data key

Architecture

Browser
  │
  ▼
Vercel — Next.js 16 (App Router)
  │
  ├── /api/ingredients         Epicure ingredient search (fuzzy, 1,790 items)
  ├── /api/recipes             Cosine similarity + allergen/safe-foods filter
  ├── /api/recipe-brief        Anthropic Claude Haiku taste-history reasoning → RecipeBrief (step 1)
  ├── /api/generate-recipe     Anthropic Claude recipe generation + validation (step 2)
  ├── /api/drink-pairings      Epicure beverage similarity search + allergen filter
  ├── /api/scan-ingredients    Thin proxy → fable-vision-ingredient-scanner Lambda (image compressed to JPEG ≤1200px client-side before upload)
  ├── /api/feedback            Recipe like/dislike storage and pattern retrieval
  ├── /api/substitutes         Embedding similarity + category scoring + Claude explanations
  ├── /api/macros              Claude Haiku on-demand macro estimation for existing recipes
  ├── /api/extract-ingredients Claude ingredient extraction from arbitrary recipe text
  ├── /api/insights            Ingredient insights (1h cache) — allergen-profile + global trends + taste profile
  └── /api/user/
       ├── profile             DynamoDB read/write (allergens, safe foods, ingredients)
       ├── saved-recipes       DynamoDB read/write (full recipe objects)
       └── collections         DynamoDB CRUD (GET, POST, PUT, DELETE)

DynamoDB tables
  ├── fable-users              Per-user profile (allergens, safeIngredients, safeFoodsMode,
  │                            ingredients[]{name, displayName, subtype, quantity, unit,
  │                            area, dateType, useByDate, boughtDate, addedAt},
  │                            kitchenEquipment[], darkMode, preferenceSignals[],
  │                            discoverSettings{}, visibleTabs[])
  ├── fable-saved-recipes      Saved recipes with full recipe JSON
  ├── fable-collections        Collections (userId+collectionId, name, recipeIds[], createdAt, updatedAt)
  ├── fable-feedback           Recipe feedback (userId+recipeId, liked, reasons, notes,
  │                            recipeTitle, recipeIngredients, allergenProfile, timestamp,
  │                            surveyResponse?{ingredientsHighlighted, ingredientsSkipped,
  │                              recipePositives, recipeNegatives})
  │                              │
  │                              ▼ DynamoDB Stream
  │                            AWS Lambda — fable-feedback-stream-processor
  │                              │  1. Extracts one preferenceSignal per ingredient → fable-users
  │                              └─ 2. Increments likeCount per ingredient in fable-ingredient-insights
  │                                    (liked events only; non-fatal; updates week + all-time records)
  └── fable-ingredient-insights  Aggregate trending data (allergenProfile PK + timeWindow SK)
                                  trendingIngredients[], trendingPairings[], trendingRecipeTypes[]
                                  Profiles: global, gluten-free, dairy-free, nut-free,
                                            gluten-free#dairy-free, vegan, low-fodmap

AWS Lambda
  ├── fable-feedback-stream-processor   DynamoDB Stream → preference signals + ingredient insights
  │                                      + sets needsRecompute = "true" on fable-users (GSI entry point)
  ├── fable-taste-profile-writer        EventBridge scheduled (every 6h) → GSI query → drift-aware
  │                                      preference analysis → Claude Haiku suggestion generation →
  │                                      writes tasteProfile to fable-users, clears needsRecompute
  └── fable-vision-ingredient-scanner   API Gateway POST /scan-ingredients → Claude Vision (Haiku 4.5)
                                         → fuzzy Epicure key matching → structured ingredient list
                                         (CJS · nodejs24.x · 30s timeout)

Shared server-side libs
  ├── lib/preference-profile.ts  buildPreferenceProfile — DynamoDB query + computePreferenceProfile
  │                              + survey merge + aggregateFormatSignals; called by /api/generate-recipe,
  │                              /api/recipe-brief, and /api/insights. Returns scores, preferred,
  │                              avoided, formatSignals.
  ├── lib/flavour-territory.ts   deriveFlavourTerritory — cosine-similarity neighbour overlap
  │                              for taste-space anchor ingredients; called by /api/insights
  │                              and /api/recipe-brief
  └── lib/survey-signals.ts      formatSignalsToClauses — maps aggregated signal keys to Claude
                                  prompt clauses via RECIPE_SIGNAL_MAP

In-memory (loaded at server startup)
  ├── Epicure Core embeddings  1,790 × 300 float32 — cosine similarity search
  └── Allergen truth table     1,790 ingredient classifications — O(1) lookup

Roadmap

✅ Completed

✅ Allergen truth table with edge cases (oat milk, almond milk etc)
✅ Custom allergen search across all 1,790 ingredients
✅ Safe Foods Mode for MCAS and restricted diets
✅ Kitchen areas (fridge, freezer, cupboard, pantry)
✅ Use-by and bought date tracking with shelf life calculation
✅ Expiry prioritisation in recipe generation
✅ Quantities and subtypes per ingredient
✅ Use my kitchen only mode
✅ Meal type filter (Snack, Starter, Main, Dessert)
✅ Cook time filter (Quick, Medium, Slow Cook)
✅ Drink pairings via Epicure embeddings
✅ Collections feature with DynamoDB persistence
✅ Like/dislike feedback system
✅ Macros toggle (off by default, eating disorder safe)
✅ Substitutes mode with context-aware scoring
✅ Recipe adaptation from pasted recipes
✅ Diet restriction presets (Vegan, Vegetarian, Keto, Low-FODMAP)
✅ Lactose intolerance with include/exclude modes and Lactaid reminder
✅ Quick-add chips filter allergens automatically
✅ Cuisine inspiration filter — 13 cuisines + Surprise me, horizontal scroll chip row
✅ Occasion filter — Weeknight, Dinner Party, Street Food, Comfort Food, Packed Lunch, Romantic Dinner, Meal Prep, Celebration
✅ Servings stepper — scale recipe quantities for 1–12 people (default 2)
✅ Kitchen equipment — Hob, Oven, Microwave, Air Fryer, Slow Cooker, Pizza Oven, Barbecue, Instant Pot (collapsible, persisted to DynamoDB)
✅ Dark mode — Moon/Sun toggle in header and allergen settings, persisted to DynamoDB
✅ Onboarding tutorial slideshow — 5-slide overlay on first launch, re-launchable from settings
✅ DynamoDB Streams + Lambda (fable-feedback-stream-processor) — real-time preferenceSignals written to fable-users on every feedback write; deployed on nodejs24.x; 6 unit tests
✅ Guest mode indicator — persistent header badge showing save-state context; tapping opens a popover explaining browser-local persistence and the coming account system
✅ fable-ingredient-insights table — aggregate trending data by allergen profile; seeded with 14 realistic records across 7 profiles × 2 time windows
✅ Discover tab — dedicated nav tab (Compass icon) between Recipe and Substitutes; Trending for you, Trending globally, Most loved, Trending pairings; each sub-section individually toggleable in settings
✅ Tab visibility settings — hide/show individual nav tabs (incl. Discover); min 2 enforced; persisted to DynamoDB. Supersedes the earlier plan to consolidate to 4 tabs: user-controlled visibility is a more flexible solution than hardcoding a merged "Explore" tab
✅ Lambda extended — liked feedback events now also write to fable-ingredient-insights (non-fatal); allergenProfile stored per feedback record
✅ /api/insights route — 1-hour cached; returns profile + global trending data
✅ 477 passing tests across 21 test suites
✅ Responsive navigation — fixed 220 px left sidebar on desktop (≥ 768 px) with Fable wordmark; bottom tab bar on mobile; same active-state and theming at both breakpoints
✅ Feedback survey — optional 4-section chip panel after every thumbs up/down; PATCH /api/feedback persists surveyResponse (all four fields: ingredientsHighlighted, ingredientsSkipped, recipePositives, recipeNegatives); ingredient signals weighted 1.5×; recipe format signals threshold-gated at 2+ appearances; 18 new tests (495 total across 22 suites)
✅ Photo ingredient recognition — camera icon in kitchen tab; Claude Vision (Haiku 4.5) via fable-vision-ingredient-scanner Lambda identifies ingredients and infers storage area; fuzzy Epicure key matching with confidence flagging; full review screen with area editing, uncertain badges, and duplicate deselection; Sonner toast notifications; 17 new tests (512 total across 23 suites)
✅ Personal taste profile + personalised discover chips — buildPreferenceProfile shared utility (DynamoDB query + preference scoring + survey merge in one call); deriveFlavourTerritory for embedding-space flavour neighbours; taste profile card on Discover (loved · avoided · flavour territory · your preferences) shown at ≥ 5 signals; Trending for you chips carry seedIngredients (top 3 liked ingredient keys) injected as a soft prompt hint into recipe generation; 20 new tests (532 total across 25 suites)
✅ Format signal injection — recipePositives + recipeNegatives from survey now flow through buildPreferenceProfile → aggregateFormatSignals (2+ appearance threshold) → formatSignalsToClauses → Claude prompt; same signals surface in the taste profile card as "Your preferences" chips with neutral display labels via SIGNAL_DISPLAY_LABELS; 5 new tests (538 total across 25 suites)
✅ Agentic two-step recipe generation — /api/recipe-brief (Claude Haiku 4.5) reasons over taste history and returns a RecipeBrief (dish direction, reasoning, novelty note, loading hints); /api/generate-recipe (Claude Sonnet) receives the brief as creative direction; brief fetch and Epicure pairings run in parallel; brief card replaces the loading spinner during recipe generation; falls back gracefully to guest hints on error or insufficient history; 26 new tests (564 total across 27 suites)
✅ Agentic taste profile evolution — fable-taste-profile-writer Lambda (EventBridge, every 6h) queries the needsRecompute-lastComputedAt-index GSI for eligible users, runs computeDriftAwareProfile (all-time vs. recent-10 diff for emerging/fading signals), calls Claude Haiku to generate 2-3 proactive recipe direction suggestions, and writes a StoredTasteProfile to fable-users; fable-feedback-stream-processor now sets needsRecompute = "true" and initialises lastComputedAt on every feedback write; /api/insights reads the stored profile when fresh (skips live buildPreferenceProfile call) and returns recipeSuggestions; Discover tab surfaces suggestions as tappable direction cards — tapping one skips the /api/recipe-brief call and uses the pre-computed direction directly; 11 new Jest tests + 12 Lambda tests (575 Jest + 12 node:test)

In Progress

Near Term

Onboarding state in DynamoDB — onboardingComplete flag currently lives in localStorage only. Add onboardingComplete: boolean to fable-users schema for authenticated users, so tutorial state persists across devices. Migration path: on auth, write onboardingComplete: false only if the localStorage flag is absent.
Nutritional database integration — USDA FoodData Central for accurate macros
Barcode/QR scanning — scan food products, auto-populate kitchen via Open Food Facts API
Ingredient substitutes improvements — better functional category matching
Equipment-aware ingredient substitution — when a recipe step requires equipment the user doesn't have, use Epicure similarity search to suggest alternative ingredients that achieve the same result with available equipment (e.g. slow cooker → hob-friendly cuts)

Medium Term

Health platform integration — Garmin Connect, Apple HealthKit, Google Health for activity-aware suggestions
Recipe cost calculator — grocery API integration (Tesco, Sainsbury's, Kroger)
Push notifications — expiring ingredient alerts
Recipe sharing — share generated recipes with friends
Native mobile app — iOS and Android for camera/barcode features

Research & Future

User authentication — Clerk or NextAuth for cross-device persistence, replacing anonymous UUID system. Guest mode remains fully functional.
Epicure Chem integration — chemical compound layer for cross-reactivity research
On-device recipe brief — migrate the /api/recipe-brief Haiku call to Liquid AI LFM2.5 running on-device; eliminates the brief round-trip entirely and keeps taste reasoning private
Editable brief direction — show the direction field as an editable text input after the brief card appears; user can nudge it ("try something spicier") before generation fires, then the brief is re-sent as the updated creative direction
Multi-turn brief refinement — "try something spicier" or "go vegetarian instead" after seeing the brief updates the direction and regenerates; brief card persists and animates to the new direction
Medical nutrition database — elemental formulas for severe MCAS
Multilingual UI — Epicure supports 7 languages
Garmin/Apple Health/Google Health integration for glucose-aware suggestions for diabetic users
High histamine preset for MCAS

Upcoming Engineering Work

AWS & Database Enhancements

1. Close the feedback loop ✅ Done

/api/generate-recipe reads the last 20 feedback records, computes ingredient preference scores, injects top 5 preferred and top 5 avoided into the Claude prompt. Auto-swap replaces avoided ingredients (score < −0.3) with the nearest Epicure cosine neighbour.

2. DynamoDB Streams + AWS Lambda — real-time preference learning ✅ Done

fable-feedback-stream-processor Lambda (lambda/feedback-processor/) fires on every write to the fable-feedback Stream, extracts one preferenceSignal per ingredient, and appends it to fable-users.preferenceSignals[] via list_append. Non-fatal on partial batch failure. 6 unit tests passing. Deployed on nodejs24.x.

3. `fable-ingredient-insights` table — aggregate analytics layer ✅ Done

fable-ingredient-insights (allergenProfile PK, timeWindow SK) is written to by the Lambda on every liked feedback event. Tracks trending ingredients, pairings, and recipe types per allergen profile. Surfaced in the new Discover section above the generation flow. Seeded with 14 records across 7 profiles.

4. TTL on recipe history ✅ Done

TTL enabled on fable-saved-recipes in AWS console. Unsaved recipes expire after 90 days (7,776,000 seconds). Saved recipes never expire. Backwards-compatible — old records without the field are treated as saved.

5. AWS Lambda — Claude Vision ingredient recognition ✅ Done

fable-vision-ingredient-scanner Lambda (lambda/vision-scanner/) accepts a base64 image via API Gateway HTTP POST, calls Claude Vision (Haiku 4.5) to identify ingredients and infer storage area, fuzzy-matches results against all 1,790 Epicure keys, and returns a structured ingredient list with confidence flags. Full review screen in the app before anything lands in the kitchen. Deployed on nodejs24.x in eu-west-2.

Key Architecture Decisions

Feedback as a personalisation engine — fable-feedback is not a log. It is the input to a preference model that shapes every subsequent Claude prompt for that user. Like/dislike + reason tags → weighted constraints → better generation over time.

DynamoDB Streams for event-driven preference updates — preference profiles in fable-users are updated reactively via Lambda on every feedback write, not on request. No latency cost at generation time.

TTL for data lifecycle hygiene — unsaved recipe history entries carry a TTL. Deliberate data expiry is not an afterthought; it is part of the table design.

Lambda as compute boundary — Claude Vision calls and stream processors run in Lambda, not in Next.js API routes. fable-vision-ingredient-scanner handles the heavy Vision call in isolation; fable-feedback-stream-processor handles the stream. Keeps serverless functions lean and gives each concern its own scaling profile. The Next.js proxy route (/api/scan-ingredients) keeps the Lambda URL server-side so it can be rotated without a frontend deploy.

API Gateway as Lambda entry point for Vision — fable-vision-ingredient-scanner is invoked via an HTTP API Gateway endpoint rather than a Function URL. Keeps the invocation pattern consistent with standard AWS architecture, allows route-level configuration, and means the Lambda URL is rotatable without any frontend changes.

Impact

250 million+ people worldwide live with food allergies
MCAS affects an estimated 17% of the population, many with severely restricted diets
575 passing automated tests across 28 suites (+ 12 Lambda tests) ensuring allergen safety and filter accuracy
Existing recipe apps are built for abundance — Fable is built for restriction
Safe Foods Mode is the only known consumer recipe tool that constrains generation to a user-defined safe ingredient list, with server-side validation to catch anything the model adds outside it
Lactose intolerance include/exclude modes with medication reminders
Macros are off by default — a deliberate decision for eating disorder recovery users

Built with Epicure (Kaikaku AI), AWS DynamoDB, Anthropic Claude, and Vercel.
H0 Hackathon submission — June 2026

Name		Name	Last commit message	Last commit date
Latest commit History 140 Commits
app		app
components		components
data		data
hooks		hooks
lambda		lambda
lib		lib
public		public
scripts		scripts
styles		styles
.gitignore		.gitignore
README.md		README.md
components.json		components.json
jest.config.js		jest.config.js
next-env.d.ts		next-env.d.ts
next.config.mjs		next.config.mjs
package.json		package.json
pnpm-lock.yaml		pnpm-lock.yaml
pnpm-workspace.yaml		pnpm-workspace.yaml
postcss.config.mjs		postcss.config.mjs
tsconfig.json		tsconfig.json
tsconfig.tsbuildinfo		tsconfig.tsbuildinfo
vercel.json		vercel.json

Folders and files

Latest commit

History

Repository files navigation

Fable — Allergen-Aware Recipe Discovery

Tech Stack

Local Development

Environment Variables

Shipped Features

Allergen System

Ingredient Input

Photo Ingredient Recognition

Recipe Generation

Drink Pairings

Recipe Feedback

Safe Foods Mode

Dark Mode

Nutritional Information (Macros)

Collections

Substitutes Mode

Diet & Lifestyle Presets

Onboarding Tutorial

Navigation & History

Discover Tab

Architecture

Roadmap

✅ Completed

In Progress

Near Term

Medium Term

Research & Future

Upcoming Engineering Work

AWS & Database Enhancements

1. Close the feedback loop ✅ Done

2. DynamoDB Streams + AWS Lambda — real-time preference learning ✅ Done

3. fable-ingredient-insights table — aggregate analytics layer ✅ Done

4. TTL on recipe history ✅ Done

5. AWS Lambda — Claude Vision ingredient recognition ✅ Done

Key Architecture Decisions

Impact

About

Resources

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Contributors

Uh oh!

Languages

3. `fable-ingredient-insights` table — aggregate analytics layer ✅ Done

Packages