Skip to main content
FromNine
Menu

AI & data

Generative AI built to keep working after go-live.

We develop generative AI applications on top of your own documents and data. Retrieval respects who may see what, every release must pass an agreed evaluation, and you can follow cost, quality and response time at all times.

For Government entities, licence holders and Maltese companies that want generative AI they can explain to the MDIA, the IDPC or the MFSA.

Start a conversationExplore enterprise knowledge
AI Engineering & GenAI: the layers we buildFour layers shown one behind the other. Top to bottom: retrieval from your own sources, the gateway that selects a model, evaluation with guardrails, and the application your staff use. A single route leads from the sources up to the application.01Retrieval from your sources02Gateway and model routing03Guardrails and evaluation04Live application
  1. 01Every answer names its source, and the system declines when it finds no evidence
  2. 02A golden test set and regression gates agreed with your staff before launch
  3. 03Routing and budgets per model, so the cost and response time of each request stay visible

01 Common problems

Why generative AI projects stall

Choosing a model is seldom the difficult part. Most programmes get stuck on what surrounds it.

  1. 01

    A convincing pilot, and no way to prove it was any good

    A handful of people tried a handful of questions and called it quality. Without a test set, each new prompt or model version is a bet.

  2. 02

    Staff see documents they have no right to see

    Content went into the index without its access rights. Your security officer and data protection officer halt the rollout, and they are correct to do so.

  3. 03

    Spending rises faster than use

    All requests go to the biggest model with the longest context. No one knows what a single answer costs, so no one can control it.

  4. 04

    Answers that sound correct and are wrong

    Retrieval brings back passages that nearly match, the model fills the gaps smoothly, and after three confident errors your staff stop trusting it.

02 Delivery scope

Our delivery scope

  1. 01 Retrieval and knowledge systems for your organisation

    We design retrieval separately for each source: how its content is parsed and split, how keyword and vector search work together, and how the access rights of each passage follow it. This is the base of an enterprise knowledge assistant.

    Our work

    • Inventory of sources, with owner, format and access model for each
    • Splitting and metadata rules for every document type
    • Keyword and vector search combined, then re-ranked
    • Retrieval that carries over the access rights of each source

    What you get

    • Ingestion pipelines that detect changes in the sources
    • A retrieval service that cites passages and knows when to abstain
    • A retrieval quality report for every source

  2. 02 Choosing models without being tied to one

    A gateway sits between your applications and the models. Through it you route each task to a suitable model, switch to a fallback when a provider is down, and change models when prices or capabilities move.

    Our work

    • Options ranging from hosted frontier models to deployments in EU regions and open-weight models you host yourself
    • Routing according to task, data sensitivity and cost
    • Timeouts, fallbacks, caching and rate limits
    • Data processing terms checked with your legal and security staff

    What you get

    • A model gateway whose routing policy lives in configuration
    • A recorded model decision for each use case
    • A budget for cost and response time per type of request

  3. 03 Evaluation treated as engineering work

    We agree quality criteria with the process owners and turn them into test sets that run whenever something changes. Checks graded by a model are compared with human review first, and only then trusted.

    Our work

    • Golden test sets for each use case, compiled with subject experts
    • Separate offline evaluation of retrieval and of generation
    • Red-team testing for prompt injection and data leaks
    • Online evaluation using production feedback and sampled reviews

    What you get

    • An evaluation suite in CI that acts as the release gate
    • Quality criteria approved by the business owner
    • Red-team results with the mitigations taken

  4. 04 LLMOps and observability

    A trace follows each request from the question to the passages found, the model call and the answer. Prompts and models carry versions in the same way as code, and budgets are enforced in production.

    Our work

    • Structured logs and traces for each request
    • Versions for prompts, models and indexes
    • Exact and semantic caching where it does no harm
    • Monitoring of quality and drift, with alerts

    What you get

    • Dashboards showing quality, cost and response time
    • A playbook for incidents specific to AI
    • A runbook for updating models and indexes

  5. 05 Governance built into the design

    The EU AI Act shapes our design choices, from classifying risk and establishing whether you act as provider or deployer, to logging, human oversight and documentation. In Malta the MDIA leads market surveillance under the Act. The legal conclusions remain with your own advisers.

    Our work

    • Classification of each use case and the obligations attached to your role
    • Technical input to GDPR data protection impact assessments
    • Design of logging and human oversight
    • System and model documentation stored next to the code

    What you get

    • A documentation pack for the system
    • A design for logging and retention
    • An oversight procedure per use case

  6. 06 Fine-tuning and document understanding, only when the tests call for it

    Fine-tuning, distillation and multimodal reading of documents close particular gaps. We turn to them once the evaluation suite proves that retrieval and good prompts fall short, never earlier.

    Our work

    • Analysis of gaps found in evaluation results
    • A plan for preparing and labelling data
    • Distillation into smaller models to save cost or time
    • Extraction that understands layout, for scans and complex forms

    What you get

    • Evaluation compared before and after the change
    • Lineage of the training data
    • A decision to retain, roll back or retire the model

03 Architecture

A reference architecture, shown for illustration

The generative AI systems we deliver for organisations usually follow one pattern: retrieval from your sources, a gateway in front of the models, guardrails before and after the model, and beneath them a loop of evaluation and observability that decides what is allowed into production.

We place the permission filter in retrieval itself, not in the screen the user sees. When someone has no right to open a document in its source system, none of its passages reach the model for that person.

Layers shown in the diagram

Sources
Wikis, document stores and business applications, together with the identity provider that records who may read what.
Retrieval
Ingestion, the hybrid index, re-ranking and the access rights filter.
Models
A gateway choosing between hosted, EU-region and self-hosted models, with caching and fallbacks.
Application
Answers with citations, a refusal when the evidence is absent, and guardrails on what goes in and out.
Operate and improve
Golden sets, regression gates, traces and budgets. Traces from production become the next test set.
Illustrative example

  1. Sources

    • Wikis and document stores
    • Business applications
    • Identity provider

  2. Retrieval

    • Ingestion and splitting
    • Hybrid search index
    • Access rights filter
    • Retrieval with re-ranking

  3. Models

    • Gateway to the models
    • Hosted, EU-region and self-hosted models
    • Guardrails on input and output

  4. Application

    • Application that cites sources

  5. Operate and improve

    • Regression gates and golden sets
    • Traces and budgets for cost and time
Retrieval, gateway and the evaluation loop

Shown for illustration only; this is not a client system.

Read the diagram as text

The main route leads from the document stores through ingestion, a hybrid index, retrieval with re-ranking and the model gateway, and ends at the application.

Business systems feed ingestion as well. The identity provider supplies a permission filter that limits retrieval. Guardrails check what goes into and comes out of the application. The gateway sends requests to hosted frontier models, deployments in EU regions or open-weight models you host.

Below sit an evaluation suite with golden sets and regression gates, and observability with traces and budgets for cost and response time. Traces from production flow back into the evaluation suite.

04 Choices and limits

Engineering choices and their limits

  • We measure quality instead of promising it

    Our approach

    Quality criteria are agreed with the owners of your processes, and the AI output is tested against them before and after every release.

    We evaluate retrieval apart from generation, so each failure leads straight to its cause.

    Limits and what we rely on

    An AI system will sometimes be wrong. For that reason we design review steps for doubtful cases and let the system abstain when it lacks evidence.

    A test set is only as strong as the experts who compile it, and we need their time at the start.

  • Access rights move with the content

    Our approach

    We record access rights when content is ingested and check them against your identity provider every time someone asks a question.

    Limits and what we rely on

    When a source system stores rights in a way we cannot read with confidence, its content stays outside the scope until that changes.

    Content shared too widely at the source stays shared too widely. Retrieval only makes it easier to find, so tidying up the source can come first.

  • Cost and response time shape the design

    Our approach

    Every type of request has a budget. We tune routing, caching and the size of the context to stay within it, and the dashboards show where the spending goes.

    Limits and what we rely on

    Providers change their prices and rate limits without asking you. The gateway gives you room to respond, but a price rise does not go away.

  • Models change beneath the application

    Our approach

    Each model is pinned to a version, and no upgrade reaches your staff before the complete evaluation suite has passed.

    Limits and what we rely on

    Providers sometimes withdraw hosted models with little warning. We keep a fallback for every use case, and the owner has to approve it.

05 Our method

Our method

Sequence matters to us: we settle how success will be measured before we settle what to build.

  1. 01

    Define the use case

    Who will use it, which decisions it supports, which sources it may read, and the cost of a wrong answer.

    ResultA use-case brief with its risk classification

  2. 02

    Compile the golden set

    Genuine questions with their expected answers, gathered with your subject experts, including those the system must refuse.

    ResultA test set under version control

  3. 03

    Get retrieval right first

    Ingestion, index and access rights filter, adjusted until retrieval reaches the quality level agreed with you.

    ResultA report on retrieval quality

  4. 04

    Release through gates

    Generation, guardrails and the gateway go into production only after the evaluation suite passes in CI.

    ResultA release backed by evaluation results

  5. 05

    Run and improve

    Traces, sampled reviews and feedback from users shape the following test set and the following release.

    ResultA monthly review of quality, cost and response time

06 People in control

Where people keep the decision

Generative AI finds information, writes drafts and makes suggestions. Your people decide what it may read, what good looks like and what happens when it is in doubt.

  • Owners approve quality

    For each use case the business owner, not only our team, approves the quality criteria and the evidence for every release.

  • Doubtful answers go to staff

    When confidence falls below a threshold you agreed, the system declines to answer or passes the question to a named team.

  • Access is decided at the source

    Your source systems and identity provider decide who may read a document. The AI layer never makes that decision.

  • Feedback improves the system

    Users can flag an answer. A person reviews each flag, and every confirmed failure is added as a test case.

07 Technologies

Technologies in our toolset

Retrieval
  • OpenSearch and Elasticsearch
  • PostgreSQL with pgvector
  • Azure AI Search
  • Cross-encoder re-rankers
Models
  • Hosted frontier models
  • Managed model services in EU regions
  • Open-weight models served with vLLM
  • Embedding models chosen per language
Evaluation and operations
  • Tracing with OpenTelemetry
  • Evaluation suites running in CI
  • Registries for prompts and models
  • Langfuse or a comparable tool
Application
  • Services in TypeScript and Python
  • Streaming APIs
  • Your current portals and the Odoo, SAP or Salesforce front end

We list a technology because our engineers have worked with it. The list does not mean that its vendor partners with or endorses FromNine.

08 Illustrative case

An illustrative case

Illustrative example

A guidance assistant over circulars for a Government department

01Starting point
Officers in a Government department answer the public from a large body of circulars, procedures and Subsidiary Legislation. Older circulars are often superseded, and two officers can give two different answers to the same question.
02The system we would put in place
Retrieval across circulars, internal procedures and the relevant Legal Notices, with access rights taken from the department’s document store and every answer pointing to the text that currently applies.
03Where staff decide
The owner of each procedure approves its test questions and signs off any release that changes the answers in that area.
04How we would judge success
How often the cited text is the one currently in force, how often the system rightly declines questions outside its scope, and the cost of each answer.

09 By sector

Sector by sector

  • Government & Public Entities

    Assistants for public officers and residents that point to the circular or the Chapter of the Laws of Malta behind each answer, and log which passages were retrieved for whom.

  • Financial Services

    Assistants on policies and procedures for MFSA licence holders, with tight entitlements, complete traceability and model risk documents your second line can check.

  • Healthcare & Life Sciences

    Retrieval over SOPs and product files for clinics and pharmaceutical plants, within firm limits: no clinical decisions, and sensitive data kept in approved environments.

Malta

Maltese authorities already designated, one test standard

Malta did not wait to organise AI Act supervision. Legal Notice 226 of 2025 made the Malta Digital Innovation Authority (MDIA) the lead market surveillance authority, single point of contact and sandbox authority, while Legal Notice 227 of 2025 gave the Information and Data Protection Commissioner (IDPC) certain Annex III high-risk areas, among them law enforcement, migration and justice.

We take this into account from discovery onwards: test sets for each language your sources use, data location set per use case, and logs and documentation that answer the questions of the MDIA, the IDPC or the MFSA without building a separate system for each. Projects co-funded under Digital Europe tend to ask for this kind of documented, reusable design.

Common questions on AI engineering

Which model would you advise us to use?

That follows from the task, the data concerned and where the system may be hosted. Most of the time we mix models: a strong model for steps that need reasoning, smaller or self-hosted ones for extraction and classification, and deployments in EU regions where the data must be kept within the EU.

Because the gateway keeps the choice reversible, our advice can change as prices and models change.

What accuracy can we expect?

We give no accuracy figure until we have measured it on your own data. Together we set quality criteria, compile a test set and report the results against it before launch and after every change.

Every model makes errors, so each design states what happens then: the system abstains, a review step catches it, or the question goes to a person.

Could our data end up training a model for somebody else?

Not with the architectures we design. We choose model services whose contract rules out training on your inputs, or we host the models in your own environment, and we record that choice for each use case together with your security staff.

Is our use case covered by the EU AI Act?

We help you work out which obligations could apply, given the use case and whether you are the provider or the deployer. We then design logging, oversight and documentation to fit, so that your legal advisers have the facts to decide. In Malta, Legal Notices 226 and 227 of 2025 made the MDIA the lead market surveillance authority and gave the IDPC certain high-risk areas.

How do you protect security and sensitive data?

FromNine holds ISO/IEC 27001 certification. We agree access boundaries, data location and retention before building starts, and the release process we put in place includes red-team tests for prompt injection and data leaks.

Where should we begin?

With a single use case that has a clear owner, a small number of sources and a test set compiled with your experts. You end up with a working system and with evidence on quality, cost and response time, which is exactly what your next investment decision requires.

Bring us one use case and the questions it struggles with

We will explain how we would test it before any talk of building it.

Start a conversationExplore enterprise knowledge