AAFAS: AI Financial Adviser

How it fits together

How it flows

1. 1

   Get the transactions

   CSV · PDF · TrueLayer

   Upload a file or connect a bank once; a stored refresh token keeps later runs going without the user.

2. 2

   Run the detectors

   Pure Python

   Subscriptions, anomalies, duplicates, health score and forecast, computed deterministically in under 5 ms.

3. 3

   Explain the findings

   5 parallel Claude calls

   Each detector's output gets its own short write-up, then one synthesis call ties them together.

4. 4

   Ask questions

   Chat session

   The session holds the full subscription list and forecast, so answers cite real amounts and dates.

5. 5

   Forget, then watch

   Firestore · Guardian

   Only findings are saved. A weekly job re-checks the account and emails when something actually changes.

My contribution

1. 01My part: the Claude orchestrator with tool use, anomaly and duplicate detection, the chat session, login and registration, and dashboard pagination.
2. 02Five pure-Python detectors with no I/O and no AI: recurring subscriptions, anomalies, duplicate services, a health score and a 30-day forecast. They run in under 5 ms, so they're cheap to run twice per session.
3. 03Recurring detection survives real data: charges are grouped per merchant, filtered by amount consistency, collapsed to one charge per month, then classified monthly or yearly by median gap.
4. 04Anomaly scoring adjusted for income: the baseline uses salary credits, so someone with low median spend doesn't get every normal purchase flagged.
5. 05Duplicate detection with per-category confidence: two VPNs are almost never intentional (0.95), two mobile plans often are (0.50), and a provider switch is detected and scored down.
6. 06Five Claude sub-agents run in parallel, one per detector, then a synthesis call becomes the chat session, so the whole analysis costs about one round-trip instead of fifteen seconds of chaining.
7. 07Chat grounded in the user's own numbers: the full subscription list and 30-day forecast go into the opening message, so "what's charging me on the 14th?" gets a real answer.
8. 08Three input paths converge on one transaction type: CSV, PDF statements parsed by Claude, and live bank data over TrueLayer OAuth.
9. 09A weekly Guardian job re-runs the detectors without the user present and only calls Claude when something fires: a new subscription, a health-score drop or an annual renewal within 7 days.
10. 10Privacy enforced in code, not only in prompts: IBANs, card numbers and emails are masked before anything reaches Claude, and snapshots store findings without a single transaction.

Challenges I solved

1. 01Letting Claude analyse someone's money without letting it make up numbers.

   Solution · Claude runs the analysis through tools that call plain Python detectors, so every figure comes from code. The system prompt also requires each claim to trace back to a tool result.

2. 02Real bank data and sandbox data look nothing alike: one charge a month versus the same subscription repeated 90 times.

   Solution · Charges collapse to one representative per calendar month before intervals are measured, so both paths run through the same code.

3. 03A flat anomaly threshold floods users with low spend and misses users with high spend.

   Solution · The baseline is adjusted by income when salary credits are present, so "unusual" means unusual for that person.

4. 04Two services in one category aren't always a mistake.

   Solution · Confidence is set per category, and a switching heuristic lowers it when one subscription starts as another ends, which is a provider swap rather than waste.

5. 05Chaining five AI calls after the detectors would make every analysis feel slow.

   Solution · The work is arranged in four phases so detectors and Claude calls overlap, bringing the wall-clock cost down to roughly one round-trip.

6. 06Money in floating-point numbers drifts.

   Solution · Every amount is whole cents in an integer, and decimals appear only when formatting for display.

What I learned

• ✦Tool use: giving a model functions to call instead of asking it to calculate.
• ✦That a privacy rule, taken seriously, becomes an architecture: with no transaction storage allowed, there is simply no schema to hold one.
• ✦Running independent AI calls in parallel, and arranging the work so the slow parts overlap.
• ✦Writing detectors as pure functions, which made them fast, deterministic and easy to unit-test.
• ✦Working milestone by milestone in a shared repository with a teammate.