FM Foundry
Reach out

We build custom foundation models for financial services to deliver exponential P&L benefits in one shot

FM FoundryReach out to us
RevolutMarket Implementation
+65%
fraud detection accuracy

-40%false positive declines

Retail BankSiriusAI- Implementation
3.5%
higher customer approval rates

80%less feature engineering effort

AffirmMarket Implementation
+3.4%
incremental approval rate

-12%30-day delinquency

Large FintechSiriusAI- Implementation
24%
collections performance improvement

21%better collections performance than existing models

Capital OneMarket Implementation
+26%
early stress detection lift

3.2xfaster model onboarding

Payments ProcessorSiriusAI- Implementation
40%
lift in customer engagement, on the same model

$70Mnannual business value generation

StripeMarket Implementation
59→97%
fraud detection for card-testing

3.8%–7%increase in authorization acceptance rates

Force Multiplier in decision systems:One AI model trained on all your rich internal data,which every function or domain can adopt for its own decisions

Traditional Decision Systems for past 20 years

One machine learning model per task

Structured Data tables

  • Model ACredit Scoring
  • Model BFraud Detection
  • Model CCollections
  • Model DMarketing

Foundation Model: new age AI decisioning brain

ONE model, adapted for many tasks and uses

Your data

  • Transactions & paymentsSwipes, merchants, amounts
  • Accounts & balancesInflows, balances, timing
  • App & service activityLogins, sessions, journeys
  • Bureau & external dataBureau and market signals

One foundation model

Foundation Model

Trained once on billions of events without losing any granularity

Unique Embeddings Profile of every customer

Every team gets access to the new found richness

  • Credit modelCredit & lending decisions
  • Fraud modelFraud & financial crime decisions
  • Collections modelCollections & recovery decisions
Billions of events, in sequenceFull customer context in every decision modelNo details lost to aggregation

For 20 years, traditional decision systems have missedthe richness of sequence and timing in your data.

Foundation models capture every detail behind every decision.

What a foundation model sees

Every event over the same 90 days, in order. The order and the spacing are the signal.

Month 1Month 2Month 3

Money in

Money out

App and digital

Service contact

Credit and risk

  1. 1

    Day 35

    Salary 3 days late, then 7 days late the next month.

  2. 2

    Days 56–63

    Balance transfer, savings order cancelled, a credit search elsewhere.

  3. 3

    Day 75 on

    Logins go from weekly to daily, mostly late-night balance checks.

  4. 4

    Day 88

    A direct debit bounces: the first thing the scorecard sees.

Behaviour profile: early financial stressHold the increase and offer early supportWarning signs caught from day 35, 53 days earlier

Foundation Model saves 70% of the technology cost and time.Traditional ML Models are built from the ground up, every time.

Traditional ML Models

0 wks

of feature work

Foundation model

0 wks

once

Traditional ML Models

Model 1Model 2Model 3
Features
Features
Features

Foundation model

Model
1
2
3
Time saved
  • Data, training and validation
  • Feature engineering
  • Events prepared once
  • Foundation model training
  • Use case

FM Foundry has proven impact across multiple clients implementations.

Custom foundation models, trained on each client's own data inside its environment, now run live decisions at leading payments and banking firms.

Recent Implementation Live
  • Foundation model

    Replaced 20-year-old decision technology in one step

  • Credit riskLive

    20% lift in performance, 4% more customers approved

  • CollectionsLive

    +25% roll-rate transition performance, lower losses

  • Cross-sellLive

    40% lift in customer engagement, on the same model

Every new use case70% less build time

10B+ data points for model training

Trained on 100M+ customer records

Every new use case80% less data prep

Solution to build your custom foundation model in 12 weeks

FM Foundry

01Data & Behavior
Readiness ProfilerEvent Ontology
02Foundation Model Factory
Tokenization & SequenceArchitecture WorkbenchModel Training Engine
03Proof & Economics
Benchmark & AblationModel Economics
04Intelligence & Inferencing
Embeddings & Task HeadsServing / Monitoring
9 modules in 4 layers
  1. Wk 1–2
    Data discovery & readiness
  2. Wk 3–4
    Event sequencing & tokenization
  3. Wk 5–8
    Foundation model, trained from the ground up
  4. Wk 8Your custom foundation model
  5. Wk 9–12
    Decision models on FM embeddings
  6. Wk 121 use cases ready for live evaluation

Frequently asked questions

Within the first quarter, compared with the models you run today.

An LLM learns language. A transaction FM learns behaviour from your own events and gives every use case a rich profile of each customer.

No. It plugs into your data platform and decision systems, and can start by strengthening the models you already run.

No. Training and scoring run inside your environment, on your credentials.

You do: the foundation model, every use case layer and the documentation.

Decisioning of tomorrow.
Available today.

Reach out to usMore about Sirius AI →
FM Foundry - Custom Foundation Models for Financial Services