Patent 40 / ML Account Management
01 / 11 US12131378B2
↑↓ navigate  ·  all patents →
Siten Sanghvi  ·  Granted Oct 29, 2024

ML Account Management

A machine learning platform that continuously monitors account activity to derive behavioral patterns, detects deviations from those patterns, identifies the anticipated transaction the deviation implies, and automatically executes the transaction or sends the user an alert — without waiting for the user to initiate the action.

US12131378B2Patent
Oct 16, 2023Filed
12 monthsTime to grant
20 Claims / 3 independentScope
0 CitationsForward citations
SCROLL TO EXPLORE
Visual patent explainer
02 / The Problem

Account management is reactive — the user must notice what needs doing and initiate every action manually.

Financial accounts generate continuous streams of activity data — balances, transaction patterns, inflows and outflows — that collectively encode behavioral patterns with predictive power. But existing account management systems don't analyze these patterns to anticipate what the user is about to need. They wait for the user to log in, identify an issue or needed action, and execute it manually. The intelligence in the data is unused.

Pattern BlindnessAccount systems store transaction history but don't derive behavioral patterns from it — the monthly auto-transfer the user always makes is not anticipated or offered proactively
Deviation Goes UnnoticedWhen activity deviates from a user's normal pattern — an unusual balance drop, a missed recurring transaction, a spending spike — no automatic system identifies this as an action signal
No Autonomous ActionEven when the right action is obvious from the pattern (low balance before a predictable large outflow), the system waits for the user to notice and act — often too late
03 / The Invention

Pattern learned → deviation detected → anticipated transaction identified → auto-execute or alert.

The platform continuously derives a behavioral pattern from the user's historical account activity. It monitors ongoing activity against this pattern. When a deviation is detected, it identifies what the anticipated transaction is — what action the deviation implies the user will need or has already triggered — and either automatically executes that transaction (if within configured thresholds) or generates and sends an alert to the user.

The dual-path response — auto-execute vs. alert — is a key feature: the platform can be configured to act autonomously for routine anticipated transactions (scheduled transfers, balance top-ups) while escalating to alert mode for higher-value or less-certain deviations. The 20 claims cover both response paths and the full deviation-to-action pipeline.

04 / Architecture

History → pattern model → live monitoring → deviation signal → identification → action.

The platform runs a continuous loop: the pattern model is updated as new activity arrives, the monitoring layer compares incoming activity against the current model, and the deviation detector signals when activity falls outside expected parameters. The anticipated transaction identifier classifies the deviation and determines the appropriate response path.

The pattern model is per-user and continuously refined — it learns from each new activity event, adjusting the expected ranges for balance levels, transaction frequencies, and recurring amounts. This self-updating model prevents the system from alerting on legitimate behavior changes (new salary, new subscription) that represent a genuine shift in the user's financial pattern rather than a deviation from it.

Architecture — US12131378B2
Historical activity
→ pattern model
Live activity
monitoring
Deviation from
pattern detected
Anticipated transaction
identified
Auto-execute
(if configured)
|
Alert user
(escalate)
05 / Pattern Derivation

The pattern model captures the structure of the user's account behavior — not just averages.

The ML pattern derivation layer analyzes historical account activity to build a behavioral model that characterizes the user's financial rhythm. This includes recurring transaction amounts and timing (the rent on the 1st, the paycheck on the 15th), typical balance ranges by day of the month, spending velocity by category, and the relationships between inflows and outflows that define this user's financial lifecycle.

The model outputs a structured pattern — not a single average balance number, but a dynamic expected-value surface that varies by time, event, and category. A deviation is identified not by distance from a global average but by departure from what the model predicts for this specific moment in the user's financial cycle.

Pattern Dimensions — US12131378B2

Recurring Transactions

Fixed-amount, fixed-timing transactions (rent, subscriptions, loan payments). Model encodes expected amount, expected date range, and expected source/destination.

Balance Rhythms

Expected balance range by day of month — low before payday, high after. Deviation at day 12 has different significance than the same deviation at day 28.

Category Velocity

Expected spend rate per category per week. A 3x spike in dining spend might be business travel (pattern-consistent) or an anomaly — model distinguishes via context.

Inflow-Outflow Relationship

The temporal relationship between income events and spending — how quickly after paycheck the user typically disperses to savings, bills, and discretionary spend.

06 / Deviation Detection

Deviation is detected against the expected pattern — not a global threshold.

The monitoring layer compares incoming account activity against the current pattern model's expected values for the current moment. A deviation is flagged when activity falls outside the expected range — balance lower than predicted, a recurring transaction that hasn't appeared by its expected date, a transaction amount significantly outside the usual range for that category.

The deviation signal carries context: it identifies which pattern element was violated, by how much, and what the violation suggests. A missing recurring transaction on its expected date suggests the anticipated transaction is the recurring payment that needs to be executed. A balance drop beyond expected range suggests a large outflow is imminent or has already occurred.

Deviation Types → Actions — US12131378B2
Deviation
Recurring transaction missing by 2+ days past expected date
Response
Anticipated transaction: execute scheduled transfer. Auto-execute if amount within threshold.
Deviation
Balance below predicted floor for this day of month
Response
Anticipated transaction: transfer from savings to cover predicted outflow. Alert user with auto-execute option.
Deviation
Category spend 3x expected velocity this week
Response
Anticipated transaction: flag for user review. Generate alert with category breakdown.
07 / Auto-Execute vs Alert

The platform acts autonomously for routine transactions — escalates to alert for high-value or uncertain ones.

Once the anticipated transaction is identified, the platform selects a response path based on the transaction type, amount, confidence level, and user-configured thresholds. Routine anticipated transactions with high confidence and amounts within threshold are executed automatically — the user doesn't need to log in and do what the system already knows is needed. High-value or lower-confidence anticipated transactions generate an alert to the user's device, presenting the recommended action for user confirmation.

The auto-execute path includes a final confirmation loop even for autonomous execution: the platform logs what was executed and notifies the user post-execution. The user can review and reverse auto-executed transactions within a configured window. This creates an autonomous-but-auditable management model rather than a purely black-box one.

Response Path Decision — US12131378B2

Auto-Execute Criteria

High confidence pattern match, amount within user-configured threshold, transaction type is routine (recurring transfer, balance top-up), no anomalous signals in session context.

Alert Criteria

Lower confidence match, amount above threshold, unusual transaction type, or first-time pattern trigger. Alert includes the anticipated transaction details and a one-tap confirm option.

Post-Execute Notification

Auto-executed transactions generate a notification to the user's device — not to request approval, but to confirm what was done and provide a reversal option within the configured window.

Threshold Configuration

Users configure auto-execute thresholds by transaction type and amount. Higher trust for small recurring transfers, explicit approval required for anything above a set amount.

08 / Pattern Refinement

The pattern model updates continuously — distinguishing genuine behavioral change from deviation.

After each deviation-and-response cycle, the platform updates the pattern model with the outcome. If the user confirmed the auto-suggested action, the pattern is reinforced. If the user dismissed the alert or reversed the auto-executed transaction, the pattern model is adjusted — preventing the same false positive from recurring. Over time, the model's deviation detection becomes more calibrated to the specific user's actual financial behavior.

The platform also detects when a deviation represents a genuine behavioral shift rather than a one-time anomaly: a new recurring subscription that appears three months in a row updates the expected-transaction model to include it. A salary change that shifts the user's inflow amount updates the balance rhythm model. The platform tracks these shifts and adjusts rather than continuing to flag them as deviations.

Model Refinement Loop — US12131378B2
Deviation detected
→ response taken
User confirms or
dismisses / reverses
Pattern model
updated with outcome
Improved deviation
detection next cycle
09 / Applications

Accounts that manage themselves around the user's financial rhythm — not the other way around.

ML-based autonomous account management transforms financial accounts from passive record-keepers into active agents that anticipate what the user needs and act accordingly — reducing overdrafts, missed payments, and balance surprises without requiring constant user attention.

Use Cases — US12131378B2
Express
Overdraft Prevention Platform detects balance will fall below the expected floor before payday. Anticipated transaction: transfer from savings. Amount within threshold → auto-executes transfer. User receives confirmation notification. Overdraft avoided without user action.
Express
Missed Recurring Payment Alert Platform detects a recurring bill payment that hasn't executed by its expected date (deviation from pattern). Anticipated transaction: recurring payment. Generates alert with one-tap confirm. User confirms → payment executed immediately.
Inferred
Spending Velocity Intervention Platform detects spend velocity 3x expected in dining category. Anticipated transaction: continued overspend against budget. Alert generated with category breakdown and suggested transfer to spending reserve. User reviews and confirms or dismisses.
Inferred
Life Event Pattern Adaptation User gets a raise — inflow amount shifts. Platform detects persistent positive deviation in balance pattern over 3 months. Updates pattern model to reflect new normal. Recalibrates all deviation thresholds to the new financial baseline.
10 / Citations

0 Forward Citations

No forward citations on record as of June 2026. US12131378B2 was granted in October 2024 — forward citations are expected to accumulate as practitioners in the autonomous financial management and ML fintech sectors reference the claims.

Forward citations confirmed via Google Patents · Jun 2026
Citation Status — US12131378B2
No citations yet — recently granted US12131378B2 granted Oct 29, 2024 Citations typically begin accumulating 12–24 months post-grant
11 / Timeline

Patent Lifecycle

Oct 16, 2023
Filed
Continuation application filed — priority Jul 29, 2020
4 months
Feb 1, 2024
Published
Pre-grant publication US20240037655A1
9 months
Oct 29, 2024
Granted
US12131378B2 granted — 12 months from filing
~19 years
~Oct 2043
Expires
Est. expiration (subject to maintenance fees)
End / Patent 40