Industry · Banking
A transformer encoder pre-trained self-supervised on tens of millions of banking event histories, downstream heads fine-tuned on the frozen embedding per institution, and specialized agents that take the score and execute — through branch, call center, and digital — inside the window the customer is still inside.
01 · SharpAI
Predictive models
11 models · foundational + downstream
02 · AUTONOMY
Agents
21 agent patterns in this vertical
03 · FDE
Engineering services
3 anonymized engagements
01 · SharpAI · Banking
One transformer-based foundation pre-trained self-supervised across tens of millions of accounts. Downstream heads attach to the frozen embedding and fine-tune to each institution's products, channels, and risk appetite — no foundation re-training per bank, no per-task data starvation.
Foundation · what the backbone tokenizes
Tens of millions of accounts, billions of events. Each dot is a raw signal type the Customer Behavior Transformer tokenizes into the customer's sequence — hover or tap any signal to see what it is.
Downstream · what the head consumes
The premium credit-card up-sell head reads the frozen CBT embedding plus a focused set of decision-time features. Hover any node — these are the inputs that drive a single score.
Foundation models · self-supervised pre-training, frozen backbone reused downstream
The shared backbone for retail and commercial banking. Tokenizes a customer's full event history — transactions, balance moves, channel touches, product holdings, life-stage transitions — into a sequence the encoder reads. The output is a customer embedding every downstream head consumes.
Transformer encoder · multi-head attention over tokenized event sequences · 256-dim shared embedding · positional + temporal encoding
Pre-training objectives
Corpus · Tens of millions of retail + commercial banking accounts across NA and LATAM · 36+ months of event history per account · multi-institution joint pre-training with federated options where data can't leave
Downstream heads · fine-tuned on the frozen foundation embedding per institution and use case
Multi-product propensity over the embedding — credit card, personal loan, mortgage, overdraft — with revenue and risk weighting.
Multi-task attention head · fine-tuned per institution
Portfolio-mix-shift propensity tuned to premium-tier eligibility and retention probability.
Linear probe over frozen embedding + calibration layer
In-branch and pre-approved personal loan propensity tuned to LATAM and NA retail patterns.
Fine-tuned head · branch-context conditioned
Early-warning on payment-stress signal months before bureau triggers — uses the temporal stress patterns the backbone already encodes.
Survival head over frozen embedding · 30/90/180-day horizons
Per-customer ranking of the lowest-cost channel that still converts — collapses contact budget without losing response.
Cost-aware ranking head · uplift-corrected
Prioritizes collection effort by recoverable amount, contact cost, and channel response.
Regression head + cost-sensitive learning
Lifecycle-aware propensity for SMB commercial banking — same backbone, firm-level aggregation.
Hierarchical head · firm-level pooling
Surfaces inbound branch visit intent — pre-approved offers and propensity at the advisor desk before the conversation.
Real-time inference head · sub-100ms latency budget
Drop-off prediction across KYC, activation, and first-90-day engagement.
Multi-horizon head · funnel-stage conditioned
Behavioral fraud anomaly head — captures pattern shifts the rule engine misses by reading the backbone's next-event surprise.
Reconstruction anomaly · backbone surprise signal
+44%
Premium credit card response · top 10 NA bank
+58%
Personal loan placement · top 1 LATAM bank
+21%
Cross-sell conversion lift
−50%
Contact budget at constant lift via channel optimization
02 · AUTONOMY · Banking
Each agent below pairs a predictive score with an executed action — across branch, call center, digital, and back-office. Outcomes are anonymized but measured against each client's prior approach.
21 agent patterns · ordered by deployment maturity
Scores every customer for premium credit card propensity; routes the offer to the highest-yield channel — branch, call center, digital.
Trigger · Daily batch + balance event
↳ +44% premium response · top 10 NA bank
Surfaces inbound branch visit context to the advisor — pre-approved offers, propensity, prior interactions — before the conversation starts.
Trigger · Customer enters branch / queues
↳ +58% loan placement · top 1 LATAM bank
Early-warning on payment-stress signal; opens recovery plays across channels with cost-aware contact strategy.
Trigger · Stress-phase threshold crossed
↳ Pre-emptive collection
Decides per-customer the lowest-cost channel that still converts — collapses contact budget without losing response.
Trigger · Per offer / per customer
↳ −50% contact budget at constant lift
Drives KYC, product activation, and first-90-day engagement; surfaces drop-off risk to the relationship team.
Trigger · New account + activation milestones
↳ Time-to-revenue compression
Predicts SMB lifecycle and cross-sell readiness for commercial banking — analogous to retail consumers.
Trigger · Quarterly + transaction shift
↳ Lifecycle-aware coverage
Identifies balance-growth windows and routes the right savings or term product to the right channel.
Trigger · Balance pattern threshold
Detects dormant or under-used cards and runs reactivation plays calibrated to the customer's historical response.
Trigger · 90/180-day dormancy
Pre-generates a one-page advisor briefing for any scheduled or expected branch visit — propensity, exposure, suitable products.
Trigger · Appointment booked / branch intent
Tier-1 conversational resolution across voice and chat — balance inquiries, card actions, payment plans — escalates on policy.
Trigger · Inbound call / chat
↳ Omnichannel deflection
Drafts and schedules outbound contact across SMS, email, voice — with frequency caps per customer and per offer.
Trigger · Campaign launch / event
Prioritizes collection portfolio by recoverable value, contact cost, and channel response; assigns to internal vs. external recovery.
Trigger · Aging bucket transition
Detects hardship indicators and routes the case to a regulated hardship workflow with explainability for compliance.
Trigger · Hardship signal + threshold
Surfaces behavioral anomalies that fall outside the rule-based fraud engine; routes to fraud ops for review.
Trigger · Pattern shift + score breach
Drives KYC document completion — reads submitted documents, flags gaps, surfaces missing items to the customer and the ops team.
Trigger · Application start
Pre-qualifies inbound mortgage interest, surfaces affordability range, and routes to the right mortgage advisor.
Trigger · Mortgage interest event
Detects retail customers who cross the threshold into private banking / wealth; routes the hand-off to the wealth team with context.
Trigger · Threshold crossing (AUM, income)
Pre-checks every customer-facing recommendation against the bank's suitability and regulatory framework before it reaches a channel.
Trigger · Pre-action gate
↳ Pre-action compliance
Surfaces winning patterns across the branch network to managers; drafts coaching content for under-performers.
Trigger · Weekly performance window
Reads complaints, classifies by severity and regulatory risk, drafts first response, and routes critical cases to humans.
Trigger · Complaint inbound
Coordinates lead scoring + outreach + branch hand-off + onboarding agents across an acquisition campaign — single audit trail.
Trigger · Campaign or partnership launch
↳ Cross-function coordination
25M+
Customers scored at a top 10 NA bank
+44%
Premium card response · top 10 NA bank
+58%
Personal loan placement · top 1 LATAM bank
−50%
Contact budget at constant lift
03 · FDE · Banking
Senior engineers embedded with the institution's credit, marketing, branch, and risk teams — through scoping, model shipping, channel integration, and the operation that follows. We integrate inside the bank's perimeter, against the bank's regulator, with the bank's data lineage requirements.
Read-only connectors into core banking, card platforms, and CRM (Salesforce / proprietary) — via API, secure direct database, or signed-URL bucket transfer. Schema mapping automated where possible.
Federated learning across regional subsidiaries and joint-venture banks — foundation models train across portfolios without raw data leaving the perimeter.
SHAP, LIME and counterfactuals on every score; per-prediction rationale logged with the decision; model cards aligned to local supervisory expectations.
Outputs land where the work happens — branch advisor screens, call center routing, CRM next-best-action, mobile banking surfaces.
MLflow + model registry + Bitbucket. Drift detection, scheduled retraining, and rollback procedures tested per model. Every training run reproducible.
Pre-action checks against the bank's suitability framework, fairness rules, and regulator-imposed contact constraints — built into the deployment, not bolted on.
Anonymized engagements · same engineering team end-to-end
Top 10 North America bank
Retail bank · 25M+ customers
Premium credit card up-sell, personal-loan in-branch propensity, and channel optimization on top of a foundational credit-scoring model. Embedded engineering team across data, modeling, and channel integration.
+44% premium credit card response · +21% cross-sell conversion · −50% contact budget at constant lift.
Top 1 bank in Latin America
Universal bank · retail + commercial
In-branch personal loan placement, savings up-sell, and SMB commercial lifecycle modeling. Branch-advisor surface integrated with the bank's core systems and CRM.
+58% personal loan placement · +16% savings lift · lifecycle-aware SMB coverage.
Regional retail bank · LATAM
Mass-affluent retail bank
Default prevention, collections strategy, and channel optimization — federated training across regional subsidiaries to keep raw data inside each jurisdiction.
Pre-emptive collection plays and contact-cost compression on the recovery portfolio.
Team shape — typical FDE composition for this vertical
Embedded senior engineers, predictive models in production, and agents that execute on every score.
