Brikz is the agentic operational layer for the Brazilian capital markets. Specialized agents for FIDC, AML and Customer Life coordinate data ingestion, document understanding, eligibility, collateral verification, monitoring and reporting on a financial foundation model mesh.
Structured Credit Agent
Onboarding, eligibility, collateral and continuous monitoring for credit funds.
AML Agent
KYC, UBO resolution, transaction monitoring and regulatory reporting.
Customer Life Agent
360° client representation for primacy, NBO/NBA and contextual credit.
The difference
There are three categories of AI in financial markets today. Each solves part of the problem. Brikz is the only class that covers regulated operations end-to-end because it sees text, transactions, ownership graphs and time series in the same decision.
| Capability | Generic LLMChatGPT, raw Gemini, writing assistants | Vertical AI AgentRPA + LLM, legal agents, vertical copilots | Agentic LDMBrikz · OS for structured credit |
|---|---|---|---|
| Reads the fund regulation PDF | Yes | Yes, domain-trained | Yes, with declared schema |
| Pulls the clause relevant to a receivable | Misses financial context | Hits its trained domain | Hits and ties to the receivable |
| Cross-checks regulation against real transactions | No | Limited | Native |
| Detects circularity in the ownership graph | No | No | Native |
| Forecasts receivables aging | No | No | Native |
| Decides automatically under fund regulation | No | No | Yes, parameterized rules |
| Returns auditable dossier with model version + applied rule | No | Partial | Yes, end-to-end |
A vertical agent runs a workflow over text. An Agentic LDM runs the regulatory cycle because it sees transactions, documents, graphs and time in the same decision — not just language.
Agents
A Brikz agent combines foundation models, tools, parameterized regulation and human handoff. It receives an event, runs the analysis cycle dictated by the applicable regulation, and returns a decision with an auditable evidence dossier.
brikz/agent-fidc
Runs the CVM 175 cycle for receivables-based investment funds across eight integrated layers, with auditable evidence per decision.
Regulation: CVM 175
Operational layers: 8
Output: decision + dossier
Stage 01 · CVM 175 · diligence
Legal onboarding
KYC documentation, ownership chain, UBO and signatory powers.
Stage 02 · CVM 175 · assignment
Eligibility & assignment
Fund regulation parameterized into rules per receivable.
Stage 03 · CVM 175 · collateral
Collateral & custody
Reconciliation with registries, uniqueness and duplicate detection.
Stage 04 · CVM 175 · monitoring
Continuous monitoring
Delinquency, subordination, substitution and operational alerts.
brikz/agent-aml
Operates the AML and counter-terrorism financing cycle prescribed by CVM 50. Designed around prioritization under an inspection budget.
Regulation: CVM 50
Reporting: within 24h
Output: investigable case
Stage 01
KYC & identity
Customer, representatives, attorneys-in-fact and core documents on a single trail.
Stage 02
Ultimate beneficial owner
Ownership chain, controllers, PEP, links and inconsistencies resolved into evidence.
Stage 03
Monitoring
Transactions modeled on a directed graph. Prioritization under an inspection budget.
Stage 04
Analysis & reporting
Risk score, exception queue, compliance dossier and regulatory report support.
brikz/agent-life
Builds a continuous representation of each retail and business client from transactions, contracts, ownership links and cash-flow series.
Adherence: BCB 119 / LGPD
Refresh: streaming
Output: client embedding
Capability 01
Dynamic primacy
Detects clients gaining or losing primacy on short windows.
Capability 02
Next Best Action
Credit, investment and product triggered by observed financial life.
Capability 03
Early financial stress
Liquidity squeeze forecast over cash-flow time series.
Capability 04
Reusable embedding
Dense per-client vector consumable by scoring, credit and fraud detection.
FIDC architecture
The Structured Credit Agent executes the CVM 175 cycle across eight connected layers. Each layer has a clear responsibility, a dedicated foundation model and an intermediate output consumable by the layers above. The audit trail is built layer by layer.
Reporting and dossier
Auditable trail → regulator, fund manager and custodian
Auditable dossier generation, manager dashboards, API for legacy systems, regulatory exports.
Continuous monitoring · CVM 175 · monitoring
Delinquency, subordination, substitution and cascade
Aging and PD forecast by the Time-series FM, senior/mezz/sub waterfall monitoring, operational alerts.
Collateral and custody · CVM 175 · collateral
Existence, uniqueness, ownership and recirculation
Reconciliation with national registries. Duplicate and recirculation detection over a directed receivables graph.
Eligibility and assignment · CVM 175 · assignment
Regulation parameterized into per-receivable rules
Regulation parser, criterion application per receivable, exception queue with rationale, automated tax treatment.
Signatory powers
Representatives, limits and signing rules
Structured extraction of bylaws, minutes and powers of attorney. Per-operation power validation. Versioning over time.
Entity resolution · CVM 175 · diligence
KYC, KYB, UBO and ownership chain
Entity, ownership, UBO, links and related parties persisted in a graph. Sanctions, PEP and watchlist screening.
Document AI
Structured multi-modal extraction
Vision + layout over fund regulation, indentures, contracts, credit notes, receipts and minutes. Declared-schema output consumable by upper layers.
Ingestion
Receivables, documents, onboarding, registries and events
Native connectors for receivables feeds, regulation, Open Finance, ERPs, registries and settlement events.
All eight layers are unified by the LFDM — the Large Financial Data Model that powers each specialized foundation model. Every decision in an upper layer carries traceability down to the raw evidence in the ingestion layer.
How it works
The platform follows a unified workflow for every agent. The same foundation models power Structured Credit, AML and Customer Life — what changes is each agent's head and the regulation it operates.
01
Transactions, registries, onboarding, documents and events arrive through native connectors on BigQuery and Pub/Sub. Dedicated São Paulo region.
02
Foundation models learn the native structure of transactions, documents, ownership graphs and financial time series. Training on TPU v5p and H100 pools.
03
Specialized heads — FIDC, AML, Customer Life — trained via LoRA. Each institution gets its own adapter on Vertex AI.
04
Serverless pay-per-query inference powers dashboards, exception queues, APIs and auditable regulator dossiers.
How to engage
Brikz ships in three distinct configurations, matched to each institution's sovereignty, integration and operational requirements. In all three, the LFDM foundation models remain Brikz IP and the per-tenant adapter remains the institution's IP.
Mode A
Brikz hosts the full infrastructure in the dedicated São Paulo region. Client accesses through web dashboard and API, configures regulation and exceptions, consumes decisions. No model ops, no GPU ops, no pipeline ops.
Who operates the AIBrikz
Who operates the businessClient
Client dataIsolated tenant
Time to production4 weeks
For fund managers, fiduciary administrators and custodians who need immediate operations without investing in an AI team.
Mode B
Client integrates the Brikz agent API inside its own product, without the Brikz dashboard. For fintechs and originators who want regulatory decision as a capability of their own product.
Who operates the AIBrikz
Who operates the productClient
IntegrationREST · webhooks
Avg integration time2 to 6 weeks
For credit fintechs, BaaS, securitizers and originators with their own digital product.
Mode C
Brikz deploys the full stack into the client's own Google Cloud project, via Terraform and Helm. Data, model and inference stay inside the client perimeter. Brikz maintains the model lifecycle via auditable runbook.
Who operates the AIBrikz · runbook
Where it runsClient GCP
Client dataNever leaves tenant
Time to production8 to 12 weeks
For mid-size banks, large fund managers and institutions with internal requirements for a dedicated tenant of their own.
Intellectual property
LFDM foundation models are Brikz IP. The LoRA adapter trained on the institution's data is the client's own IP and stays cryptographically isolated.
Data residency
In every delivery mode, client data stays in the São Paulo region with Confidential Compute, in-transit and at-rest encryption, and customer-managed keys (CMEK) on request.
Platform
The Large Financial Data Model (LFDM) is the composition of four specialized encoders, each covering a native modality of financial data. Agents consume embeddings from these encoders through trainable heads.
Modality
Multi-modal structured extraction over fund regulation, indentures, contracts and credit notes.
vision + layout
Modality
Aging, PD and cash-flow forecasting over long windows with Mamba-2 backbone.
temporal
Modality
Contextual scoring on tables of originators, payors and market counterparties.
tabular · TabPFN-v2 tuned
Modality
Inference on directed transaction and ownership graphs. Circularity detection.
graph · GraphSAGE production
Infrastructure
Brikz runs 100% on Google Cloud in a dedicated São Paulo region, with Gemini and Vertex AI as the foundation of the product. The choice solves four requirements at once: regulated data sovereignty, TPU access for foundation-model pretraining, first-class graph support in both the warehouse and the operational store, and serverless pay-per-query inference.
Google Cloud Partner
Brikz is part of the Google for Startups Cloud Program and acts as a co-development partner on Vertex AI, Gemini, BigQuery Graph and Spanner Graph applied to regulated financial services.
Foundation models
Gemini · Vertex AI Model Garden
Gemini Pro and Flash families via Vertex AI Model Garden, orchestrated as the reasoning engine for the agents.
Agentic workspace
Gemini Enterprise
Brikz agents with data residency, IAM controls and audit logging from the enterprise tier of Gemini.
Compute · AI
Vertex AI
Training, fine-tuning, MLOps and serving for foundation models and per-tenant LoRA adapters.
Acceleration
TPU v5p / Trillium
Foundation model pretraining on TPU pools. Complementary support for H100/H200 GPUs.
Warehouse
BigQuery
Serverless lakehouse, pay-per-query, with native embeddings via BQML for SQL-native ML.
Analytical graph
BigQuery Graph
Graph queries over transactional history and ownership chain inside the warehouse.
Operational graph
Spanner Graph
Operational representation of the originator/payor/UBO graph with strong consistency.
Document AI
Document AI
Structured multi-modal extraction over fund regulation, indentures, contracts and minutes.
Serving
Cloud Run · GKE
Serverless inference and always-warm regional endpoints for production agents.
Streaming
Pub/Sub · Dataflow
Real-time ingestion of payments, settlement events and registry webhooks.
Storage
Cloud Storage
Datalake in open Iceberg format, partitioned by LoadDate, with auditable lineage.
Isolation
Confidential Compute
Cryptographic isolation of execution environment per regulated tenant.
Governance
Dataplex · IAM
Data catalog, lineage and granular access control for internal and external audit.
Region
São Paulo
Data sovereignty in dedicated Brazil region. Adherent to LGPD and Central Bank Res. 119.
Scale
Petabyte-scale processing with LoadDate partitioning over Cloud Storage and BigQuery.
Inference
Always-warm endpoints on Cloud Run. GraphSAGE at 740K edges/second.
Training
TPU v5p and H100/H200 GPUs on the same platform plane. Per-tenant LoRA on Vertex AI.
Sovereignty
Data in dedicated region. Confidential Compute for regulated tenant isolation.
Adaptation
LoRA, QLoRA and DoRA per-tenant in the Vertex AI Model Registry.
Audit
Rationale, model version, adapter version and applied regulation snapshot on every execution.
Additional stack
Expansion roadmap
Every vertical to the right reuses the same foundation layer. The documentary, regulatory and transactional complexity of structured credit is the technical floor — adjacent segments consume the same engine with additional heads.
Current
Fund managers, fiduciary administrators, custodians and structurers operating under CVM 175 and CVM 50.
brikz/agent-fidc · agent-aml
Next
CRA, CRI, incentivized debentures, securitizers and venture-debt funds on the same documentary and collateral pipeline.
head extension
Expansion
Credit originators, SCDs, BaaS and acquirers — reuse of the AML and Customer Life agents for scale.
agent-life focus
Horizon
Mid-size and digital banks modernizing compliance and credit on top of Brazilian Open Finance.
platform expansion
Company
Brikz Tecnologia Ltda is a Brazilian limited liability company, created to build the agentic operating layer of the national capital markets. Early stage, backed by institutional capital and financial market professionals.
Security and compliance
The platform adopts controls expected by Brazilian regulated financial institutions from day one, with a roadmap for international certifications.
LGPD
Personal data handled in adherence to the Brazilian General Data Protection Law, with documented legal basis per operation.
Central Bank Res. 119
Cyber risk management and cloud processing contracting compliant with Brazilian Central Bank Resolution 119.
Confidential Compute
Cryptographic isolation of the execution environment per regulated tenant on Google Cloud.
ISO 27001 · SOC 2 Type II
International information security management certifications planned during the first operational year.
Observed metrics
Numbers below are measured on AMLSim with 124M synthetic transactions, on NVIDIA A100 80GB on Vertex AI, using the standard AML Agent pipeline. Reproducible in an isolated Brikz environment on request.
Ranking
87.7%
Recall@1% — GraphSAGE concentrates illicit transactions in the top-1% prefix.
Efficiency
7.94
Inspections per illicit (IPI@1%) — operational metric under inspection budget.
Throughput
740K/s
Edges scored per second on NVIDIA A100 80GB.
Test volume
124M
Transactions in the AML1M scenario — comparable to months of payments at mid-bank scale.
Documentation
Brikz's technical documentation covers data connectors, model lifecycle, agent specifications, infrastructure on Google Cloud and governance. Same documentation consumed by client engineering and compliance teams in production.
Open documentation →Start
Introduction to LFDM
concept · architecture · scope
Agents
Technical specification
FIDC · AML · Customer Life
Data
Connectors & datasets
CNAB · registries · payments
Models
Training & adaptation
Vertex AI · TPU · LoRA
Serving
Inference & API
Cloud Run · vLLM · latency
Security
Governance & audit
LGPD · BCB 119 · trail
Brikz Labs
Standalone pages with synthetic data. Each demo replicates a single operational layer of the LFDM at reduced scale, renders directly in the browser, and generates fresh data on every run. Technical coverage spans AML graph prioritization, ownership chain resolution, PD and aging forecasting, regulation parsing, zero-shot tabular scoring, and community detection over cession networks.
Open Brikz Labs →Demo 01 · AML
Transaction prioritization
GraphSAGE · Recall@k
Demo 02 · Graph FM
Ownership chain
UBO · cycle · PEP
Demo 03 · Time-series
PD and aging forecast
Mamba-2 · stress
Demo 04 · Document FM
Regulation parser
CVM 175 · SQL
Demo 05 · Tabular FM
Zero-shot scoring
TabICLv2 · in-context
Demo 06 · Graph FM
Network communities
GraphAny · systemic risk
For technology, compliance, risk and product teams at regulated financial institutions in Brazil.
contato@brikz.io · São Paulo, Brazil · Built on Google Cloud