× STEPS
Case study | Sovereign LLM, Afreximbank

Llama 3, an LLM fine-tuned on the bank's data

For Afreximbank, STEPS fine-tuned Llama 3 on the bank’s internal data: corpus building, training, evaluation and deployment of an open source model the bank stays in control of.

Open source model
Internal data
Secure deployment
01 - The context

Generative AI, but not at any price

The bank wanted to use generative AI on its own documentation: reports, policies, internal notes. But off-the-shelf solutions raised confidentiality and relevance issues.

!

Sensitive documentsA bank's internal data cannot be entrusted to an external service without guarantees.

!

Specialised vocabularyGeneral-purpose models know little about trade finance and the bank's own language.

!

A dependency to avoidBuilding on a proprietary API means depending on its prices, rules and availability.

02 - The choice

Why a model of its own rather than an API

A bank cannot send its internal documents just anywhere. A fine-tuned open source model stays under its control.

GENERAL-PURPOSE LLM VIA API

Convenient, but

  • Data sent to a third party
  • Business vocabulary poorly handled
  • Dependency on a vendor
FINE-TUNED LLAMA 3

The bank's choice

  • Model hosted and controlled by the bank
  • The bank's language and context learned
  • Guardrails defined in-house
03 - The approach

From corpus to model in production

01Corpus

Collection, cleaning and anonymisation of internal documents.

02Fine-tuning

Training Llama 3 on this corpus, in the bank's vocabulary.

03Evaluation

Answer quality tests and guardrails before any go-live.

04Go-live

Secure deployment, controlled access and production monitoring.

04 - The uses

Querying the bank's documents

The model powers an information retrieval platform: teams upload their documents and ask questions in natural language, with built-in feedback.

05 - The benefits

What the model brings to the bank

An AI asset the bank owns, controls and can evolve.

Main benefitSovereignty over its AI

The bank has a language model tailored to its business, hosted and controlled by it: its data never leaves and it decides how the model evolves.

ConfidentialityProtected data

The corpus and the model stay in a controlled environment.

RelevanceBusiness-aware answers

The model understands the bank's vocabulary and context.

CostsNo pay-per-use billing

An open source model avoids paying a vendor for every request.

IndependenceNo vendor lock-in

The model can be retrained, moved or replaced freely.

ProductivityFaster search

Teams find information in documents using natural language.

FoundationA base for other uses

The same model powers assistants, summaries and AI agents.

06 - Our work

Four phases, from idea to model in production

PHASE 01Scoping

Choice of use cases, documents to include and confidentiality rules.

PHASE 02Data

Collection, cleaning, anonymisation and structuring of the training corpus.

PHASE 03Training

Llama 3 fine-tuning, iterations and evaluation on business questions.

PHASE 04Production

Secure deployment, search platform and user feedback loop.

07 - Architecture

Under the hood

A complete MLOps chain, from data preparation to production monitoring.

Llama 3Fine-tuningLoRAPythonRAGMLOps
  • DataInternal corpus cleaned, deduplicated and anonymised
  • ModelOpen source Llama 3, fine-tuned on the bank's corpus
  • EvaluationBusiness question sets and safety guardrails
  • ServingHosted in an environment controlled by the bank
  • InterfaceInformation retrieval platform with user feedback
FAQ

Frequently asked questions

Why Llama 3?

It is a high-performing open source model: the bank can host it, adapt it and evolve it without depending on a vendor.

Does the data leave the bank?

No. The corpus, training and serving stay in an environment controlled by the bank.

Does training need to be redone regularly?

The model can be retrained when the corpus evolves; the approach is documented and reproducible.

Want generative AI that stays in-house? Let's talk about your sovereign LLM.

Book a scoping session ↗