Crayon Data Crayon Data Tangram AI
AI Marketplace · Fully agentic · ✷
BANKING · CUSTOMER EXPERIENCE & VALUE SUITE
A I   M A R K E T P L A C E

Ranks 65M+ merchant genomes into one personalised decision in under 1.5s -
turning segment-blasted offers into individually relevant ones, live at HDFC.

<1.5s ranked decision ✦ 37.8% visit-to-view ✦ $600M+ incremental spend enabled
Why AI Marketplace exists

One offer for everyone. Zero relevance measured.

The problem

Relevance assumed, unmeasured

Every customer sees the same blanket offer, and no one measures whether it actually lands.

  • 4.8% baseline visit-to-view under blanket segment offers
  • 53% of customers would switch banks over impersonal service
  • ~50% of marketing budget spent on offers, unranked to the individual
Today, how it operates

Static segments, one message

A marketer builds one segment, writes one message and waits weeks to build the next.

  • Weeks to stand up a single new segment campaign
  • One-size offer per segment, regardless of individual signal
  • No loop back from offer shown to outcome measured
With AI Marketplace

Every customer, their own ranking

The engine ranks every offer against each customer's own signal, in real time.

  • 65M+ genomes ranked into each decision
  • <1.5s from signal to a personalised choice
  • 37.8% visit-to-view, up from the 4.8% baseline
What it does · three moves

Standardize it. Score it. Personalize it.

01 Merchant Standardization Turn raw strings into one merchant
  • AI-clusters name variants at 85–96% similarity
  • 20 AI-generated clusters, ranked by transaction impact
  • Approved with one click, ready for enrichment
02 Taste Graph Score what the spend really says
  • Jaccard, cosine or weighted-Jaccard affinity scoring
  • Merchant-to-merchant co-occurrence, e.g. Spotify–Netflix 0.91
  • Spend clustered into named tastes, e.g. “Food & Transport”
03 Choice Generation Rank the next best offer, per customer
  • Category, cross-category, discovery & next-best lists
  • Confidence and transaction thresholds, tuned live
  • Every customer, their own ranked list
The capability set

Six things it does, from raw data to ranked choice.

One pipeline, not six disconnected tools — each stage feeds the next automatically.

01 Any-source ingestion

CSV upload, S3 bucket, or direct DB connect.

02 Schema mapping

Raw columns, mapped to one standard model.

03 AI merchant standardization

Embeddings + clustering, human-approved.

04 Merchant enrichment Human input

Websites, tags and categories, auto-identified.

05 Taste graph scoring

Affinity between merchants, not just categories.

06 Real-time choice generation

Category, discovery & next-best, ranked per customer.

Dashboard

Every visit, scored continuously.

live at HDFC

  1. 1
    6.1M visits tracked +12.5% over the prior period, updated live
  2. 2
    87.99K claims, 28.8% claims/views Every offer view traced through to outcome
  3. 3
    94% data cleanliness, 5.0% CTR Both moving up, both watched continuously
1 2 3
Data Studio · Onboard

Any source. One studio.

CSV, S3, or direct DB

  1. 1
    Upload, connect, or query CSV drag-and-drop, an S3 bucket, or Postgres/MySQL/Snowflake
  2. 2
    1.5M rows, ready in one pass Transactions_Q3_Raw onboarded and marked ready
  3. 3
    Every dataset, tracked Row and column counts logged the moment data lands
1 2 3
Data Studio · Schema Mapper

Six raw columns. One standard model.

Map once, reuse everywhere

  1. 1
    Mapped automatically cust_id_raw → customer_id, txn_amount → amount, and so on
  2. 2
    Human-confirmed, not guessed Suggested mappings, approved with a click
  3. 3
    One schema downstream Every source lands in the same standardized shape
1 2 3
Merchant Enrichment

1,184 of 1,240 merchants, already identified.

95.5% identification rate

  1. 1
    1,240 new canonical merchants Created straight from standardization
  2. 2
    1,120 ready for enrichment 90.3% ready, tags and categories auto-extracted
  3. 3
    56 flagged for a human Missing websites routed for manual input, not guessed
1 2 3
Recommendation Core · Feature Store

Thirty features. Recomputed daily.

Customer, merchant & customer-category

  1. 1
    10 + 10 + 10 features Customer, merchant, and customer-category, side by side
  2. 2
    Built for ranking, not reporting Recency, frequency, spend, churn risk, lifetime value
  3. 3
    Every feature, explained Min / max / mean / std-dev shown against every column — the inputs the Taste Graph scores next
1 2 3
Recommendation Core · Taste Graph

Fifteen connections. One affinity score each.

Jaccard, cosine or weighted-Jaccard

  1. 1
    15 connections, 0.66 average affinity Three similarity metrics on tap — Jaccard, cosine, or weighted-Jaccard
  2. 2
    Spotify–Netflix, 0.91 Merchant-to-merchant co-occurrence, ranked strong to moderate
  3. 3
    Spend clustered into named tastes “Food & Transport” is the top cluster today
1 2 3
Recommendation Core · Choice Generation

150 choices a customer. 94% covered.

Category, cross-category, discovery & next-best

  1. 1
    5 lists × 30 choices, 150 total Category, cross-category, discovery, repeat and next-best lists, per customer
  2. 2
    94% coverage, 90% avg confidence Confidence and transaction thresholds, tuned live
  3. 3
    Three cohorts, one engine Data Rich, Data Dark and Inactive customers each get their own strategy
1 2 3
In summary

Every customer, their own ranking.

It ranks 0M+ merchant genomes into one personalised
decision in under 0s, lifts visit-to-view to 0%
— and has already unlocked 0M+ in incremental spend, live at HDFC.

It is live at HDFC Bank today, and it can be live on your platform next.

Try it first on your own transaction data in the hosted sandbox.

demo.crayondata.ai/ai-marketplace

AI Marketplace · Crayon Data

Thank you.

Scan for the live portal

QR code to the live portal