Each colour maps a published value to the exact source passage used to support it. Only bounded excerpts are public; administrators can inspect the complete captured source.
Title
SumatoSoft builds a private LLM compliance copilot inside a European digital bank's own cloud
SumatoSoft built a private large language model copilot that runs inside the bank's own cloud, grounds every answer in the bank's policies, and cites the clause behind it.
Description
A licensed European digital bank (Lithuania) had compliance analysts reading policy by hand and writing case narratives from long transaction histories. SumatoSoft built a private RAG copilot with a LoRA fine-tuned Llama 3 model running in the bank's Azure account, citing the policy clause behind each answer, with analyst sign-off. The source reports onboarding reviews of about five business days now take about two, and case narratives about 90 minutes down to about 25.
Complex onboarding reviews that used to take about five business days now take about two,
Writing a case narrative dropped from about 90 minutes to about 25
Company
SumatoSoft
SumatoSoft built a private large language model copilot
Country
Lithuania
Industries: Region: Fintech, digital banking, and consumer lending Lithuania PROJECT TYPE Custom LLM development, production build (retrieval- augmented gen…
Industry
Financial Services
A licensed European digital bank offers accounts, cards, consumer loans, and payments across the EU.
Problem
As its customer base grew, the bank's compliance team spent most of its day reading policy by hand and writing cases from long transaction histories. Transaction monitoring generated a high volume of alerts, the large majority of them false positives, and onboarding customers who needed enhanced due diligence meant reading across several long policy documents that changed often.
As its customer base grew, the compliance team spent most of its day reading policy by hand and writing cases from long transaction histories.
Transaction monitoring generated a high volume of alerts, the large majority of them false positives, and each one took an analyst's time to clear.
Solution
SumatoSoft built a private large language model copilot as a service within the bank's Azure cloud account, with AWS Bedrock as an alternative hosting option. A retrieval-augmented generation pipeline indexes the bank's AML, KYC, sanctions and fraud policies and a history of resolved cases, with hybrid search over PostgreSQL with pgvector and Qdrant; every answer names the clause it drew on and the copilot abstains when retrieval finds nothing that supports an answer. An open-source Llama 3 model fine-tuned with LoRA on the bank's own casework drafts alert briefings and case narratives, agentic steps handle repetitive work such as summarizing transaction histories and grouping alerts, and a human analyst signs off on everything before it leaves the system.
SumatoSoft built the copilot as a private service within the bank's Azure cloud account, with AWS Bedrock as an alternative hosting option.
SumatoSoft fine-tuned an open-source Llama 3 model with LoRA on the bank's own casework.
A human analyst signs off on everything before it leaves the system.
Implementation approach
The project ran for 8 months. Before launch, SumatoSoft red-teamed the system and built an evaluation harness that scores output quality on every release, and the copilot now runs in production across the bank's alert and onboarding workflows.
Before launch, SumatoSoft red-teamed the system and built an evaluation harness that scores output quality on every release.
The copilot runs in production across the bank's alert and onboarding workflows.
Business value
Within the first year of rollout, the compliance team handled a larger book of business without adding headcount in proportion. According to the source, complex onboarding reviews dropped from about five business days to about two, alert triage moved from around 30 cleared alerts per analyst per day to between 70 and 80, case narratives dropped from about 90 minutes to about 25, audit requests that once took three weeks are answered in about four days, and new analysts reach full productivity in about 6 weeks rather than 3 months.
Alert triage moved from around 30 cleared alerts per analyst per day to between 70 and 80
Audit requests that once took three weeks are answered in about four days
New analysts reach full productivity in about 6 weeks rather than 3 months
AI capabilities
Generative AI, Agentic AI
a set of agentic steps handles the repetitive work: summarizing a transaction history into a briefing, grouping alerts that share a pattern, drafting the case narrative, and flagging gaps for the analyst.
Business functions
Legal & Compliance
The team indexed the bank's AML, KYC, sanctions, and fraud policies along with a history of resolved cases
Technology
Llama 3, LangChain, LlamaIndex, PostgreSQL, Qdrant
A hybrid search over PostgreSQL with pgvector and Qdrant combines keyword and semantic matching.
LangChain and LlamaIndex orchestrate the retrieval and the drafting
SumatoSoft fine-tuned an open-source Llama 3 model with LoRA on the bank's own casework.
Deployment model
cloud
SumatoSoft built the copilot as a private service within the bank's Azure cloud account
Deployment options
cloud
SumatoSoft built the copilot as a private service within the bank's Azure cloud account
Headline outcome
Complex onboarding reviews that used to take about five business days now take about two,
Use case type
Knowledge search
Cites the exact policy clause behind every answer, and abstains when the policy base does not cover the question
