Foundation · what the backbone tokenizes
Every signal the backbone reads about the debtor.
The customer's event history across accounts, cards, loans, channels and past collection interactions. Hover or tap any signal to see what it is.
Solution · Collections for banking & fintech
Most collection strategies still run on rules and deciles: everyone in the same bucket gets the same call, on the same day, with the same message. We replace that with a decision system — predictive models that know who will pay, who will pay because you acted, and how to reach them — and agents that execute the treatment across WhatsApp, voice, SMS and email, inside your policy and your regulator's rules.
01 · SharpAI
Predictive models
9 models · foundational + downstream
02 · AUTONOMY
Agents
9 agent patterns in this vertical
03 · FDE
Engineering services
2 anonymized engagements
The delinquency lifecycle
Every stage of the lifecycle has its own decision, its own cost of contact and its own window. One model per stage, one shared view of the customer behind all of them.
01
Current · pre-delinquency
Detect payment stress weeks before the due date, from the customer's own behavior — not from a bureau trigger that arrives late.
Decisions we model
02
1–30 days past due
Separate the customers who will pay on their own from the ones who need a nudge — and leave the first group alone.
Decisions we model
03
31–90 days past due
Concentrate effort where the action changes the outcome: the right channel, the right moment and the right message per customer.
Decisions we model
04
90+ days past due
Offer the plan the customer can actually keep, and decide what to work in-house and what to place with an agency.
Decisions we model
05
Every stage
Every contact, promise and payment feeds back into the models. The last conversation is not noise — it is context for the next action.
Decisions we model
01 · SharpAI · Collections
Collections heads attach to the same Customer Behavior Transformer we use across banking. The embedding already encodes how each customer earns, spends, pays and responds — so the collections models start from a rich view of the customer instead of a handful of bureau variables, and each head answers one operational question.
Foundation · what the backbone tokenizes
The customer's event history across accounts, cards, loans, channels and past collection interactions. Hover or tap any signal to see what it is.
Downstream · what the head consumes
The uplift head does not ask who is likely to pay. It asks who pays because of the action. Hover any node — these are the inputs behind one treatment decision.
Foundation models · self-supervised pre-training, frozen backbone reused downstream
The shared banking backbone. It tokenizes each customer's event history — transactions, income, balances, payments, channel touches and prior collection interactions — into a sequence the encoder reads, and produces the customer embedding every collections head consumes.
Transformer encoder over the latest customer events · 256-dim shared embedding · positional + temporal encoding
Pre-training objectives
Corpus · Fine-tuned on each institution's own portfolio inside its own cloud · pseudonymized data · works from the data available on day one and improves as the loop feeds back
Downstream heads · fine-tuned on the frozen foundation embedding per institution and use case
Probability of entering delinquency at 30, 90 and 180 days — the early-warning signal that opens the lifecycle.
Survival head over the frozen embedding
Who pays, and when, without any action — the baseline every treatment is measured against.
Calibrated classification head
Estimates how much the customer can realistically pay now, from income and spending patterns.
Regression head on cashflow features
The incremental effect of each action on each customer — so effort goes to the people it changes, not the ones who would pay anyway.
Uplift modeling · T-learner and X-learner
The channel and time window in which each customer is most likely to respond.
Ranking head over channel response history
Chooses tone, content and offer framing per customer and keeps learning from every response.
Contextual bandit
Which customers will accept a payment plan or refinancing, and which plan structure they are most likely to keep.
Propensity head + plan-structure ranking
Probability that a promise will be kept, so follow-up is scheduled where it matters.
Classification head over conversation and payment history
02 · AUTONOMY · Collections
Prediction creates insight; decision and execution create impact. A team of specialized agents, coordinated by supervisory agents, turns each score into a treatment — and every treatment runs through the institution's policy, consent and compliance rules before it reaches the customer.
9 agent patterns · ordered by deployment maturity
Monitors payment-stress signals on current accounts and opens a preventive play before the due date — a reminder, a rescheduled debit date or an offer.
Trigger · Stress score threshold
↳ Delinquency prevented, not chased
Sends the lightest effective nudge to early-stage customers likely to self-cure, and stays silent where silence works better.
Trigger · Due date / 1–30 days
↳ Lower contact cost
Ranks the daily portfolio by incremental recovery per contact and assigns each account to agents, dialer, digital or no action.
Trigger · Daily portfolio refresh
↳ Effort where it moves recovery
Holds the conversation on WhatsApp Business — approved templates, opt-in and 24-hour window enforced — explains the balance, shares payment links and answers questions.
Trigger · Treatment assigned to WhatsApp
↳ Self-service resolution
Proposes the payment plan the customer can keep, within the discount and term limits the institution has authorized.
Trigger · Renegotiation propensity
↳ Plans that are kept
Registers each promise, schedules the follow-up and adapts the next step if the payment does not arrive.
Trigger · Promise registered
↳ Higher kept rate
Hands the case to a human collector — with full context and a recommended next step — when the conversation, amount or vulnerability signal requires it.
Trigger · Policy or judgment threshold
↳ Humans on the cases that need them
Checks every action before it executes: contact hours, frequency caps, consent, do-not-contact lists, message wording and data-protection rules.
Trigger · Every action
↳ Auditable by design
Reads the outcome of every contact and feeds it back into the models and the strategy, running champion-versus-challenger tests continuously.
Trigger · Continuous
↳ A strategy that improves every cycle
03 · FDE · Collections
We do not ask you to switch off what works. We run a controlled challenger against your current strategy — on your portfolio, in your cloud, through your channels — and measure incremental recovery against a control group. This is not body-shopping. It is forward-deployed engineering.
Core banking, card processor, rule engine, CRM, dialer and collection agencies — connected read-only first. Model outputs land in the systems your team already operates, replacing fixed deciles with as many treatments as the data supports.
A control group on your current strategy and test arms that separate the effect of better models from the effect of better execution — so every point of recovery is attributable.
WhatsApp Business, voice, SMS, RCS, email and call center orchestrated from one decision layer, with frequency caps and contact windows enforced at the orchestration level.
Pseudonymized data, consent and opt-out management, contact-hour rules and a full audit trail of every decision — aligned to local consumer-protection and data-protection regulation (LGPD in Brazil, Ley 25.326 in Argentina).
Recovery rate, incremental recovery, promise-to-pay kept rate, recovery per contact and cost-to-collect — with guardrails on complaints, opt-outs and contact saturation.
Two weeks of scoping, then modeling, controlled execution and a results-backed business case. We can start with limited data and grow the scope as the loop proves itself.
Anonymized engagements · same engineering team end-to-end
Retail bank · LATAM
Loan and card portfolios
Model-driven segmentation of the collections portfolio to replace rule-based deciles with finer treatments loaded into the bank's rule engine and CRM — including detection of administrative false delinquency and multi-loan customers. Runs inside the bank's cloud.
In delivery — impact measured pre/post against the current rules.
Top 5 telco · LATAM
B2B invoice portfolio
Payment and non-payment prediction per invoice, with probability-based prioritization and best-channel strategy, validated offline on the operator's historical data.
10.7% potential collection lift identified in offline validation.
Team shape — typical FDE composition for this vertical
Embedded senior engineers, predictive models in production, and agents that execute on every score.