Skip to content
EN

Solutions · Generative and multimodal reporting

Generative and multimodal reporting

From data and information to reporting in text, voice and video.

On one page

We build systems that turn data, sources and knowledge into structured reporting, explanations, narratives and multimodal pieces adapted to different audiences, languages, durations and channels.

  1. 01From structured information to multimodal generation. Generation begins before any drafting.
  2. 03An architecture that separates information, planning and generation. Not every decision inside a generative system has to be generative.
  3. 05Architectural principle. Data, calculations, entities, context and constraints are kept separate from the linguistic realisation.

We combine information processing, calculation, NLG, generative models and validation inside an architecture where each component takes on what it can do best.

From structured information to multimodal generation.

Generation begins before any drafting. The system gathers information, normalises data, identifies entities and relationships, calculates whatever can be calculated and builds a structured representation of the available content.

From that representation it can decide what information is relevant, how to organise it, what level of detail each audience needs and what constraints it must preserve during generation.

Generative models come into that architecture as components of linguistic realisation, explanation and adaptation, not as substitutes for the data, calculation, knowledge or validation layers.

A single informational base can produce a report, a briefing, an alert, a narrative, subtitles or a video piece without rebuilding the whole process from scratch for each format.

What we build

Ingestion and normalisation of information

Integrating heterogeneous sources into a usable representation. We connect structured data, APIs, documents, events, publications and other sources. We normalise formats, entities, dates, units, relationships and context so that the rest of the system can work on a common base.

Informational representation

Separating the available information from the way it will later be communicated. We structure facts, metrics, entities, events, relationships, periods, context and metadata through typed schemas. Deterministic calculations and transformations are carried out before the linguistic generation whenever they can be resolved exactly in code.

Selection and NLG planning

Deciding what to communicate before deciding how to express it. The system selects information according to objective, audience, context and moment. It can establish priorities, depth, order, narrative relationships, length and structure before asking the model for the linguistic realisation. Content planning stays separate from drafting.

Controlled generation

Using generative models where they contribute linguistic flexibility. The models turn structures and content plans into natural language conditioned by terminology, format, audience, language, length and content constraints. The architecture avoids delegating to the model tasks that can be better resolved through structured data, logic or deterministic calculation.

Adaptation by audience and context

The same information may need different explanations. From a common representation, versions can be generated with different depth, register, language, duration, structure or level of prior knowledge. The system distinguishes between information that must be preserved and aspects of the communication that can be adapted.

Multimodal realisation

Text, voice and video as different ways of expressing a single informational base. The system can produce reports, scripts, narrations, audio, subtitles, graphics and video. Each modality introduces constraints of its own regarding duration, pacing, pronunciation, segmentation, synchronisation and correspondence between language and visual elements.

Validation and evaluation

Checking data, content and linguistic realisation in different layers. We can validate figures, dates, names, signs, units, periods, entities, relationships and structural constraints against the data the system used. Separately, we evaluate properties such as coverage, coherence, semantic fidelity, terminology, naturalness, fitness for the audience and correspondence between modalities.

Reporting architecture
  1. Sources
  2. Data and information
  3. Structured representation
  4. Selection and planning
  5. NLG
  6. Text · voice · video
  7. Validation
One informational base · many realisations
  1. Data + context
  2. Representation
  3. Planning
  4. BriefReportExplanation
  5. Languages
  6. Text · voice · video

An architecture that separates information, planning and generation.

Not every decision inside a generative system has to be generative.

Deterministic

Calculations · figures · dates · units · identifiers · exact transformations

Structured

Entities · events · metrics · relationships · context · metadata · content plans

Generative

Drafting · explanation · synthesis · reformulation · adaptation · narrative composition

Evaluated

Coverage · consistency · terminology · semantic fidelity · naturalness · multimodal fitness

One information system, multiple reporting formats

  • Automated, recurring reporting
  • Alerts and briefings based on data or events
  • Reports and executive summaries
  • Explanations generated from structured information
  • Narratives adapted by audience
  • Multilingual and localised reporting
  • Generated audio and narrations
  • Subtitles and synchronised versions
  • Video generated from a single informational base
  • Multi-format systems where text, voice and video share data, context and controls

Architectural principle

First we structure the information. Then we decide how to communicate it.

Data, calculations, entities, context and constraints are kept separate from the linguistic realisation. Generative models contribute flexibility where it is useful, while logic, structures and validators control whatever can be resolved more precisely outside the model.

One informational base. Many realisations.

Separating information from linguistic realisation makes it possible to reuse a single representation across different products and channels.

The same set of data can feed a twenty-second briefing, a long report, an explanation for a general audience, another for specialists, different languages, an audio narration and an audiovisual piece.

What changes is the way of communicating. The underlying information can remain common.

Components of the architecture

Data and information

Ingestion · APIs · normalisation · structured extraction · typed schemas · entity resolution · temporal context

Representation and knowledge

Entities · relationships · events · metrics · taxonomies · terminology · context · metadata

NLG

Content selection · document planning · conditioned generation · structured generation · templates · LLMs · constraints

Calculation and control

Deterministic transformations · rules · calculations · validation of figures · units · dates · entities · consistency

Linguistic evaluation

Semantic fidelity · coverage · entailment · terminology · coherence · criterion-based evaluation

Multimodal

TTS · ASR · subtitling · temporal alignment · synchronisation · audiovisual composition

Operation

Observability · versioning · traceability · continuous evaluation · human-in-the-loop according to risk

Generation is only one layer of the system.

The most useful reporting systems do not begin by asking a model what it should write. They begin by organising data, context, calculations and communication objectives.

From there, NLG and generative models make it possible to turn that information into explanations, reports and narratives adapted to different people, languages and media.

We design the complete chain: from the data and the information to their expression in text, voice and video.

The system from the inside

Maturity · Evaluation
How it is organised
We explicitly separate sources, informational representation, calculation, content selection, NLG planning, linguistic realisation and multimodal rendering. This separation makes it possible to change models, languages, audiences or formats without turning the whole reporting chain into a single generative call. Data and constraints travel as typed structures and can be reused by different forms of output.
What can be examined
The system can preserve the relationship between an output and the data, sources, calculations and transformations that took part in generating it. We can examine the input information, the normalised structures, the calculations, the content selection, the narrative plan, the linguistic output and the validation results. This makes it possible to analyse not only what the system produced, but also what information it used and what transformations it applied to produce it.
How it is checked
We do not treat quality as a single score. Properties that can be checked exactly — figures, signs, dates, names, units, periods, relationships or structural constraints — are validated through code where possible. Coverage, semantic fidelity, coherence, terminology, naturalness and fitness for the audience require different evaluation mechanisms. Model-based evaluators can be part of the system, but they do not replace deterministic checks where those exist.

Conditions and limits

Structured generation reduces errors and improves control, but it does not automatically make any output correct. The final quality also depends on the quality and currency of the data, on the selection rules, on the calculations performed, on the available context and on the criteria used to evaluate the output. Voice and video additionally bring constraints of their own regarding pronunciation, synchronisation, visual selection, timing and correspondence between modalities. We design these layers according to the domain, the risk and the level of autonomy each system actually needs.

A capability takes on value inside a specific problem.

Share the context with us and we will think together about how to combine information, technology, software, people and evaluation around the result that matters.

Search · Type to search. Esc to close.

Search results →

Ask the bot

This is Ask 3.14, an automatic assistant. It answers only with the public content of this website — solutions, capabilities, systems, questions, readings and laboratory work —, shows the content it has used and distinguishes what the website does not yet allow it to establish. It is not a person from the team and it does not know your case; to speak with someone, write to the team.

Automatic assistant limited to the public content of 3.14; it may abstain. To speak with a person, write to the team.

Write to the team →