What if your money coach understood your financial life, not just your transactions?
MetaMint, Independent [2026]
Overview
Problem
Financial decisions don’t happen in isolation. Planning a trip, splitting costs with friends, and staying within a budget are connected but managing money often means piecing together information across conversations and financial tools.
Solution
MetaMint is a conversational AI money coach designed to make financial guidance more contextual by connecting everyday conversations with spending and budgeting decisions.
The product has two connected layers:
- Product Vision
Money Coach
Personalized guidance for trip planning, savings goals, budget awareness, and relevant financial products, informed by the user’s financial context.
- Working Prototype
Group Expense Tracking
An expense-understanding and tracking capability that turns group conversations and receipts into structured expenses, shared balances, and planned-versus-actual spending data.
The expense-tracking layer provides a foundation for the broader Money Coach vision, connecting what people plan to spend with what they actually spend.
Users
Individuals seeking more contextual guidance on spending and budgeting, alongside friends and travel groups coordinating shared expenses.
Results
- 01
Designed
The Money Coach
Designed a conversational AI money coach to help people plan trips, set savings goals, stay within budgets, and explore relevant financial products using the context of their everyday lives.
- 02
Built
Group Expense Tracking
Built a working prototype that understands expense-related messages and receipts, splits group expenses, tracks balances, and links planned expenses to actual spending.
- 03
Evaluated
AI Reliability
- 92%accuracy (23/25 images) in image relevance classification.
- 96.6%accuracy (28/29 messages) in intent classification during the unified evaluation run.
Problem
01. Planning a trip without losing track
- Meet Elena!
Planning a birthday trip to Yosemite
She’s currently unemployed, so she needs to be careful about how much she spends without giving up the experience.
Elena keeps track of her expenses in her head, on scraps of paper, and across different conversations.
With so much to remember, it’s easy to lose track of bills or even estimate how much she is spending.
I wish it could keep track of my expenses as I plan, so I can see where my money is going and stay within my budget.
Discuss expenses in chat
Share payments, receipts, and plans across conversations.
Reconstruct expenses
Manually enter amounts, identify who paid, and calculate how to split costs.
Settle up with friends
Use a separate tool to track balances and figure out who owes whom.
Calculate personal spending
Manually add up expenses, track individual spending, and work out how much money remains for the trip.
02. The opportunity
Elena doesn’t just need another tool to split a bill. She needs a way to connect everyday spending with her bigger financial picture, so she can track what she spends, plan what she can afford, and make more informed decisions.
But also
Product Strategy and Insights
01. JTBD
To understand what users truly need from a financial companion, I mapped their functional, emotional, and social Jobs to Be Done (JTBD), looking beyond expense tracking to the broader decisions they want help making.
| # | Type | Description | Jobs to be done |
|---|---|---|---|
| 2 | Functional | Split shared expenses fairly and pick the right card for each purchase, without leaving the chat | When my group chat is planning a trip or settling a bill, I want a fast and fair way to track and split costs without leaving the chat, so I can keep things simple and accurate. |
| 3 | Emotional | Feel in control and confident about money decisions without active management or anxiety | When I’m thinking about a financial decision or upcoming expense, I want to feel confident I’m making the right call quickly, so I can stop worrying and move on with my day. |
| 4 | Social | Be the easygoing friend who handles money fairly and gracefully in groups | When I’m splitting expenses or making money decisions in a group, I want everything handled fairly without anyone getting upset, so I can be the friend who keeps things together without making it weird. |
- 2Functional
Split shared expenses fairly and pick the right card for each purchase, without leaving the chat
When my group chat is planning a trip or settling a bill, I want a fast and fair way to track and split costs without leaving the chat, so I can keep things simple and accurate.
- 3Emotional
Feel in control and confident about money decisions without active management or anxiety
When I’m thinking about a financial decision or upcoming expense, I want to feel confident I’m making the right call quickly, so I can stop worrying and move on with my day.
- 4Social
Be the easygoing friend who handles money fairly and gracefully in groups
When I’m splitting expenses or making money decisions in a group, I want everything handled fairly without anyone getting upset, so I can be the friend who keeps things together without making it weird.
03. Product Analytics
Why the product exists, who it serves, and the metrics that show whether it is working.
PRODUCT RATIONALE
Real world utility beyond social and content
Defensible new revenue stream without becoming a regulated financial entity
Demonstrate Meta’s capability to extend WhatsApp into life-utility categories, setting a template for future products
Address Meta’s brand-trust deficit in finance through the No-Authority Zone principle
COMPANY GOAL
EngagementDistribution is solved at Meta-scale; The real question is whether users engage with the AI when given the option. If group chats invoke @Mint regularly and users opt into the 1:1 personal coach, growth and monetization follow naturally.
LIFECYCLE FOCUS
- Growth
- Our focusEngagement
- Retention
- Monetisation
COMPETITORS
- Cleo
- Splitwise
- CreditKarma
- ChatGPT
HOW SUCCESS IS MEASURED
NORTH STAR METRIC
Decision Events per Active User per Month
- Split confirmed
- Card Recommendation
- Settlement completed
- Personal Coach
GROWTH METRICS
- First Meaningful Interaction (FMI) Rate
- Group Activation Rate
ENGAGEMENT METRICS
- Group Chat Sessions per Activated Group per Month
- 1:1 Activation Rate
- AI Conversation Depth
MONETISATION METRICS
- Referral Click-Through rate
- Partner Referral Revenue
Solution
01. See it in action
Money Coach
Designed the broader Money Coach experience, spanning trip planning, savings goals, spending guidance, and contextual card recommendations.
Group Expense Tracking
Built a working prototype that uses AI to classify expense-related inputs, extract receipt details, and interpret user intent, with deterministic code handling expense splitting and balance tracking.
Hover over a screen to see its video play (tap on a touch screen).
01. Equal Expense Splitting
MetaMint extracts the bill details and automatically splits the total equally among group members.
02. Itemized Expense Splitting
Members can split a bill based on individual items rather than dividing the total equally.
03. Intelligent Expense Filtering
MetaMint distinguishes relevant restaurant and grocery bills from personal images and unrelated content, ignoring inputs that shouldn't be tracked.
04. UPI Payment Recognition
MetaMint recognizes UPI payment confirmations made to businesses and converts eligible payments into equally split group expenses.
05. Personal Budget Tracking
Users can track their spending, monitor remaining budgets, and view their financial position directly in their personal chat.
06. Planned Expense Tracking
Users can mention an upcoming expense, track it as planned, and confirm it once the transaction actually happens.
02. How group expense tracking works
Each box shows what it does, then what it is built on.
- InputReceipt image
Photo or screenshot
- InputChat message
Text in the group chat
- AIImage classifier
Process or ignore
Anthropic API (vision)
- AIExtraction
Totals, line items
Anthropic API (vision)
- AIIntent classifier
5 intent types
Anthropic API (text)
- CodeNormalization
Standard expense
Python
- CodeSplit engine
Who owes whom
Python, AI item matching
- CodePlanned state
Status tracking
Python, swipe-to-reply
- CodeBudget
Spent and left
Python (budget.py)
- BackendFlask backend and pipeline
app.py and pipeline.py
Python, Flask
- UIWhatsApp-style prototype
Group chat, Me chat, tracker
HTML and JavaScript
- EvalEval harness
Scores classifier, extraction, intent
run_evals.py, labeled test folders
- AI (Anthropic API)
- Deterministic code
- Input, backend, UI
03. One key engineering decision
AI interprets. Code calculates.
THE CHALLENGE
Expense messages and receipts are messy and unpredictable, but financial calculations need to be consistent.
THE DECISION
I separated AI interpretation from deterministic financial logic. AI handles the ambiguous inputs (classifying images, extracting receipt details, and interpreting expense intent), while code handles expense normalization, bill splitting, and balance updates.
AI INTERPRETS
Messy, ambiguous inputs
- Classifying images
- Extracting receipt details
- Interpreting expense intent
CODE CALCULATES
Consistent financial logic
- Expense normalization
- Bill splitting
- Balance updates
WHY IT WORKS
This separation makes financial calculations predictable and testable without relying on an LLM to perform money math.
- Predictable
- Testable
- No LLM money math
EVALUATIONS AND RESULTS
01. Evaluating AI reliability
I built a unified evaluation script to test MetaMint’s AI components against labeled examples, covering image relevance classification and message intent classification. For receipt extraction, I reviewed outputs manually, but a labeled test set was not yet available to calculate accuracy.
Rather than optimize for accuracy alone, I examined failure cases to understand where the system could misinterpret financial activity and how those errors should be handled.
IMAGE RELEVANCE CLASSIFIER
92.0%
23 of 25 images
2 payment-confirmation images missed
INTENT CLASSIFIER
96.6%
28 of 29 messages
1 ambiguous expense update missed
- Manual review
RECEIPT EXTRACTION
Not yet scored
20 line items found on one reviewed receipt; 1 price misread
The classifier scores measure agreement with the labeled test cases, not overall real-world product accuracy.
02. One Failure, One Decision
- Case 01 · Image classifier
When a payment confirmation looks like a personal transfer
THE FAILURE
The classifier first scored 94.1% (16/17 images). After I added more UPI payment screens it fell to 86.4% (19/22), missing bare payment screens where the recipient’s name looked like a person or a code. One prompt revision brought it to 92.0% (23/25).
THE DECISION
I checked first that the misses were real, not mislabeled data. I stopped tuning when two ambiguous cases remained, because more changes risked overfitting and wrongly processing genuine personal transfers.
THE PRODUCT IMPLICATION
When the classifier is uncertain, the safer behavior is to let the user explicitly request processing rather than automatically treat every payment screen as an expense.
A restaurant payment that MetaMint did not recognize as a restaurant from its name. - Case 02 · Intent classifier
When an expense update lacks context
THE FAILURE
“The Airbnb ended up being 18k” was misclassified because the message alone didn’t make it clear whether the user was reporting a new expense or updating a planned one.
THE DECISION
I documented the limitation and relied on the reply-to-planned-expense flow to provide context when users respond directly to the original plan.
THE PRODUCT IMPLICATION
The right classification can depend on conversation context, not just the wording of a single message.
03. How the evals shaped the product
Testing the AI was a way to decide how the product behaves when the AI is wrong.
Act when invited
FINDING
Two bare UPI screens stayed ambiguous at 92%.
DECISION
I stopped tuning. If MetaMint ignores a screen, the user can ask it to process it.
Fail safe
FINDING
In a finance product, recording something nobody asked for is worse than missing it.
DECISION
Ambiguous images lean ignore. Ambiguous messages lean informational.
Test cases define the product
FINDING
“Dinner cost us 2,400, but the expense is on me.” has nothing to split.
DECISION
I relabeled it informational. An expense is something that triggers a split.
Money math stays in code
FINDING
Split, ledger, and budget amounts have one right answer.
DECISION
Plain code computes them, so the evals cover only judgment calls.
04. Limits and what I would do next
WHAT THE NUMBERS DO NOT SHOW
- Receipt extraction is not scored yet. None of the 14 receipts has expected values.
- One price was misread on a 20-item receipt ($17.99 as $7.99). I documented it as a known limitation.
- The item matcher, the newest AI call, has no eval set.
- No user testing yet.
WHAT I WOULD DO NEXT
- Label the 14 receipts to score extraction.
- Add a check that compares the sum of line items to the receipt total.
- Build an eval set for the item matcher.
- Test with real users.
Every miss on record has a known cause and a handling decision.
Learnings
- 01A missed expense beats a wrong one.
In a finance product, recording something nobody asked for does more harm than missing it, so when the model is unsure, the product stays quiet.
- 02For money, trust comes before personalization.
A coach that remembers your finances has to show what it remembers and ask first, so memory consent and a visible summary made the MVP while monetization waited.
- 03Let AI judge and code calculate.
The model handles messy inputs, while split, ledger, and budget math stay in plain code that can be tested.