Defined task
Begin with the question, user, decision boundary and evidence needed to support a useful outcome.
RESEARCH / SPECIALISATION / EVALUATION
Different missions.
Different intelligence.
Exania develops and evaluates specialised AI approaches for the information, risks and operating conditions of distinct sectors. Each model family begins with a defined problem, not a predetermined technology.
EXANIAATLASPUBLIC SYSTEMS
SENTINELCYBER
HELIOSRESEARCH
MERIDIANRISK
AEGISHEALTH
VECTORINDUSTRY
SPECIALISED SYSTEM MAP / EXANIA01 / A CLEAR DEFINITIONWHAT IS A MODEL FAMILY?
The right model is defined by what it needs to solve.
In real organisations, the difficult part is rarely choosing the largest model. It is deciding which information can be used, what a correct result looks like and how people will verify that result.
An Exania model family names a specialist area of research and system engineering. For a particular engagement, the resulting solution may combine existing foundation models, task-specific machine learning, retrieval, algorithms and proprietary software components. Architecture, validation and availability depend on the project's scope.
How the platform is engineeredRelevant data
Specify the source, quality, permissions and formats of the information the system may process.
Measurable behaviour
Evaluate performance, uncertainty, security and human review before operational use.
02 / SIX SPECIALISATIONSSELECT A MISSION AREA
Six families.
Distinct objectives.
The names organise Exania's work around the domain. They are not a claim that one underlying model is appropriate for every use case, nor a guarantee of off-the-shelf deployment.
01 / PUBLIC SYSTEMSATLAS
Turn complex institutional records into traceable, reviewable knowledge.
Explore model 02 / SECURITY OPERATIONSSENTINELCorrelate technical signals and support structured cyber investigation.
Explore model 03 / RESEARCH SYSTEMSHELIOSAccelerate scientific information review and structured hypothesis exploration.
Explore model 04 / FINANCIAL ANALYTICSMERIDIANBring structured analysis and uncertainty into financial operations.
Explore model 05 / LIFE SCIENCESAEGISOrganise complex biomedical information with rigorous expert oversight.
Explore model 06 / INDUSTRIAL SYSTEMSVECTORAnalyse operational signals and engineering knowledge for complex systems.
Explore model03 / MODEL FAMILY EXPLORERTECHNICAL PROFILE / SELECT TO CHANGE
Inside each family.
Explore the information types, analytical approaches, expected outputs and evaluation questions behind each specialisation. These are practical engineering profiles, not performance claims.
MODEL FAMILY / 01
ATLAS
Government & institutional intelligence
Clarity across institutional information.
ATLAS is the public sector and institutional research family. It is designed around large collections of policies, official records, legislation, operating procedures and other domain-specific material that teams must interpret accurately.
Help analysts locate relevant evidence, connect related records and prepare structured briefings without losing the original source, version or context.
01 / INPUT INFORMATION
- Legislation and policy documents
- Institutional repositories and reports
- Approved records, tables and correspondence
02 / TECHNICAL APPROACH
- Retrieval with source attribution
- Language understanding and classification
- Entity extraction and knowledge linking
03 / REVIEWABLE OUTPUT
- Source-linked research briefs
- Topic and document classification
- Evidence maps and review queues
SAFEGUARDS & HUMAN REVIEW
Access controls, provenance checks, retention policies and explicit human approval before consequential decisions.
HOW PERFORMANCE IS EVALUATED
Source attribution accuracy, retrieval relevance, unsupported-claim rate, completeness and policy-sensitive error analysis.
PRACTICAL SCENARIO
A public agency needs to reconcile hundreds of policy updates and internal directives. ATLAS can help identify the relevant texts, compare revisions and prepare a sourced analysis for legal or operational review.
LANGUAGE / RETRIEVAL / GRAPHDiscuss this domain
MODEL FAMILY / 02
SENTINEL
Cybersecurity & digital resilience
Make sense of a noisy threat environment.
SENTINEL is the cybersecurity and digital resilience family. It focuses on security telemetry, event records, network observations and technical intelligence that often arrive in different formats and at high volume.
Assist authorised security teams with triage, anomaly detection, evidence correlation and preparation of incident narratives for qualified investigators.
01 / INPUT INFORMATION
- Security events and approved logs
- Asset and vulnerability inventories
- Alert histories and incident records
02 / TECHNICAL APPROACH
- Time-series and anomaly modelling
- Rules with statistical correlation
- Evidence retrieval and incident classification
03 / REVIEWABLE OUTPUT
- Prioritised review queues
- Linked incident timelines
- Explainable anomaly and trend reports
SAFEGUARDS & HUMAN REVIEW
Human approval for response actions, strictly scoped access to systems, and retention rules for potentially sensitive telemetry.
HOW PERFORMANCE IS EVALUATED
False positive/negative rates, precision and recall on labelled incidents, detection latency and drift monitoring.
PRACTICAL SCENARIO
A security operations team sees overlapping alerts across authorised infrastructure. SENTINEL can group related observations, show a timeline and surface evidence that analysts should review before taking action.
ANOMALY / SEQUENCE / RULESDiscuss this domain
MODEL FAMILY / 03
HELIOS
Scientific discovery & advanced research
A research engine built around evidence.
HELIOS is the scientific research family. It explores analysis of specialist publications, experimental results, scientific imagery and technical datasets to support research teams facing complex questions.
Support literature review, extraction of findings, identification of related work and prioritisation of questions for further human-led investigation.
01 / INPUT INFORMATION
- Peer-reviewed literature and preprints
- Approved experimental datasets
- Research metadata, figures and observations
02 / TECHNICAL APPROACH
- Scientific text and semantic retrieval
- Pattern analysis and modelling
- Computational comparisons and simulations
03 / REVIEWABLE OUTPUT
- Evidence tables and literature maps
- Reproducible analysis summaries
- Research leads with uncertainty notes
SAFEGUARDS & HUMAN REVIEW
Versioned datasets, reproducibility records, scientific validation and careful separation of hypotheses from established findings.
HOW PERFORMANCE IS EVALUATED
Citation fidelity, reproducibility, retrieval coverage, uncertainty calibration and domain-expert review.
PRACTICAL SCENARIO
A laboratory needs to compare findings across years of technical publications and study records. HELIOS can organise the evidence and highlight conflicting or missing results for researchers to examine.
RESEARCH / MULTIMODAL / ANALYSISDiscuss this domain
MODEL FAMILY / 04
MERIDIAN
Financial systems & risk intelligence
Move from raw financial data to measured insight.
MERIDIAN is the financial research and risk analytics family. It combines document interpretation, numerical analysis and forecasting approaches for organisations that need structured, auditable views of changing conditions.
Support authorised teams in comparing exposures, modelling possible scenarios, detecting anomalies and assembling evidence for risk discussions.
01 / INPUT INFORMATION
- Approved financial statements and reports
- Structured time-series datasets
- Operating indicators and risk registers
02 / TECHNICAL APPROACH
- Statistical forecasting and anomaly detection
- Document analysis and retrieval
- Scenario testing and probabilistic methods
03 / REVIEWABLE OUTPUT
- Scenario comparisons and sensitivity views
- Data-quality and exception flags
- Evidence-based risk briefing materials
SAFEGUARDS & HUMAN REVIEW
Controlled access to financial information, explicit assumptions, approved data sources and qualified review before decisions.
HOW PERFORMANCE IS EVALUATED
Back-testing, stability under distribution shift, calibration, explanation quality and data lineage.
PRACTICAL SCENARIO
An enterprise risk unit compares different market or operational conditions. MERIDIAN can organise inputs, evaluate scenario assumptions and show uncertainty so specialists can review possible exposures.
FORECASTING / ANALYSIS / RETRIEVALDiscuss this domain
MODEL FAMILY / 05
AEGIS
Healthcare & life sciences research
Evidence first, where the stakes are highest.
AEGIS is the healthcare and life sciences research family. Its focus is scientific literature, biomedical data and knowledge workflows that require particularly careful treatment of quality, privacy and context.
Help qualified research and clinical-domain teams search publications, interpret approved research datasets and organise scientific evidence for expert investigation.
01 / INPUT INFORMATION
- Biomedical publications and clinical guidelines
- Approved or de-identified research datasets
- Laboratory and study records
02 / TECHNICAL APPROACH
- Biomedical language processing
- Multimodal pattern analysis
- Evidence retrieval and structured extraction
03 / REVIEWABLE OUTPUT
- Literature synthesis with references
- Research evidence tables
- Anomaly or pattern review candidates
SAFEGUARDS & HUMAN REVIEW
Data minimisation, privacy protection, subject-matter validation and no unsupervised diagnosis or treatment decisions.
HOW PERFORMANCE IS EVALUATED
Clinical-domain review, sensitivity/specificity where relevant, subgroup error analysis, source faithfulness and privacy assessment.
PRACTICAL SCENARIO
A life sciences research team is reviewing biomedical literature and study outputs. AEGIS can structure the findings and trace them to original sources for specialist assessment.
BIOMEDICAL / LANGUAGE / VISIONDiscuss this domain
MODEL FAMILY / 06
VECTOR
Industry, energy & critical infrastructure
Bring intelligence closer to industrial reality.
VECTOR is the industrial systems, energy and infrastructure family. It considers asset telemetry, technical documentation, geospatial information and operational history within defined industrial environments.
Support maintenance planning, equipment trend analysis, anomaly investigation and expert decision support across approved operational workflows.
01 / INPUT INFORMATION
- Equipment telemetry and sensor history
- Maintenance and engineering records
- Approved imagery and geospatial datasets
02 / TECHNICAL APPROACH
- Time-series and predictive modelling
- Computer vision and pattern recognition
- Rules, optimisation and simulation
03 / REVIEWABLE OUTPUT
- Asset-condition indicators and trends
- Maintenance review candidates
- Operational scenario comparisons
SAFEGUARDS & HUMAN REVIEW
Safety engineering constraints, physical process isolation, authorised operational access and human sign-off on interventions.
HOW PERFORMANCE IS EVALUATED
Failure prediction performance, alert burden, robustness, uncertainty estimates and site-specific validation.
PRACTICAL SCENARIO
An infrastructure operator wants to understand repeated equipment alarms. VECTOR can analyse authorised sensor trends and maintenance records to support a structured inspection decision.
TIME-SERIES / VISION / SIMULATIONDiscuss this domain
WORKING PRINCIPLE
Models can assist analysis, search and prediction. In consequential or high-risk settings, qualified people retain the authority to interpret evidence and approve actions.
04 / TECHNICAL FOUNDATIONSMODELS ARE MORE THAN LANGUAGE
Choose the method
the task requires.
The architecture follows the data and the evidence standard. Language models can be useful, but numerical algorithms, computer vision, statistical methods or explicit rules are often equally important.
Transformers & language models
Analyse text, classify records, retrieve relevant information and create drafts grounded in authorised sources. Source checking and human review are essential.
DOCUMENTS / LANGUAGE / RETRIEVALComputer vision
Extract features from images and video; support recognition, classification and visual pattern analysis under appropriate data and privacy controls.
IMAGERY / SPATIAL FEATURESTime series & forecasting
Study changes in ordered observations, identify anomalies and quantify possible future conditions, including the uncertainty of estimates.
SENSORS / RISK / OPERATIONAL SIGNALSHybrid and structured reasoning
Combine learned patterns with rules, search, optimisation or graph techniques when transparent constraints and reliable processes matter.
RULES / KNOWLEDGE / PLANNING05 / MODEL SELECTIONQUESTION → DATA → SYSTEM
From mission to
model design.
Rather than assign a name and begin training, Exania's model engineering approach assesses the task, information constraints and validation needs before choosing a method.
01 / DEFINEProblem & scope
What decision or analytical task needs support? Who validates the output?
02 / EXAMINEData suitability
Are sources lawful, complete, relevant, current and sufficiently representative?
03 / SELECTModel & method
Compare retrieval, language, statistical, vision and hybrid architectures.
04 / TESTEvidence & limits
Measure domain-specific quality, failure cases and operational risk.
05 / INTEGRATEControlled workflow
Deploy with access rules, audit records and human review points.
DESIGN DECISION
Capability is not the same as readiness.
A research prototype may demonstrate a useful technique without being suitable for a production environment. Operational suitability depends on the quality of its evidence, performance, security and deployment controls.
06 / VALIDATION & EVALUATIONMEASURE WHAT MATTERS
Performance must be
demonstrated.
Accuracy alone rarely describes a useful AI system. Models should be evaluated against mission-specific errors, incomplete evidence, uncertainty, privacy and operational impact.
EVALUATION AREA / 01
Task performance
Define a representative test set and measure the failures that matter to the organisation. Relevant metrics may include precision, recall, forecast error or correct classification.
Representative test dataBaseline comparisonDomain-specific error review
METHODS AND ACCEPTANCE CRITERIA ARE DETERMINED FOR EACH PROJECT.07 / GOVERNED INTELLIGENCEBUILT FOR ACCOUNTABLE USE
Private by design.
Reviewed by people.
Sensitive work requires more than an accurate prediction. Teams must define who may use the system, where information may be processed and what happens when a model is uncertain.
Access boundaries
Permissions and identity controls for the users, sources and environments involved in each workflow.
Data governance
Documented provenance, data minimisation, retention and handling requirements appropriate to the project.
Explainable outcomes
Evidence references, uncertainty indicators and clear documentation of model assumptions or limitations.
Human authority
Review and escalation mechanisms; approval of consequential outcomes remains with authorised professionals.
08 / APPLICATIONSILLUMINATING A POSSIBLE WORKFLOW
What does this look like in practice?
Select a sector to see how a specific analytical problem can map to one model family and a controlled workflow. The actual approach would depend on data access and project assessment.
Policy and public information
A ministry needs to compare updated rules and official reports. The design begins with authorised records, multilingual search and a clear audit trail.
RELEVANT FAMILYATLAS
PRIMARY OUTCOMETraceable document review
09 / TECHNICAL QUESTIONSMODELS, RESPONSIBILITY & SCOPE
Questions worth
asking first.
Model engineering begins by understanding the boundary conditions. These answers outline how Exania approaches specialisation.
Speak with the teamAre ATLAS, SENTINEL and the other families separate foundation models?
They are names for domains of specialised AI engineering. A project may use one or several architectures, including task-specific models, licensed or open foundation models, retrieval systems, numerical methods and custom software. A family name does not itself specify a training method or a production release.
Can a system use an organisation's internal information?
That is assessed project by project. The data must be authorised for the intended purpose and handled through agreed controls for identity, retention, access, processing location and review.
Does every project need to train a new model?
No. Retrieval, established algorithms, fine-tuning or other integrations may be a better technical and economic choice. We evaluate existing capabilities first and reserve additional training for problems where it is justified.
How do you know if a model is reliable enough?
Before deployment, the project needs a representative evaluation set, task-specific metrics, failure analysis and operational acceptance criteria. Monitoring and periodic re-evaluation are important as real-world data and requirements change.
Can Exania operate in private environments?
Private cloud, dedicated remote and local infrastructure are deployment patterns considered by the platform. Feasibility depends on hardware, licensing, network topology, regulatory requirements and the workload; no particular environment is guaranteed without a project assessment.
Can AI make final decisions without human review?
High-impact workflows should specify responsible human operators and explicit approval boundaries. The degree of automation is governed by the task's risk, applicable requirements and documented safeguards.
What research underpins these specialisations?
Exania's AI research interests include neural networks, natural language processing, computer vision, statistical learning, hybrid systems, planning and explainability. Some directions remain experimental and must not be interpreted as deployable products.
10 / BUILD FOR YOUR MISSIONEXANIA / EXOHOOD LABS INC.
Start with the problem.
Design the intelligence.
Share your mission, available data, decision constraints and preferred environment. We can discuss which research methods and system architecture are appropriate to evaluate.