Timeliness matters
Forecasts must reflect the data available at issuance time
Notification, consolidation and revision have different delays. Point-in-time provenance is essential for a fair prospective evaluation.
Brazil · Epidemiological intelligence · Dengue
Anticipate risk.
Plan the response.
Dengo AI turns epidemiological, climate and territorial data into municipal probabilistic forecasts for the next four weeks. Current results are retrospective; independent validation is the next gate before operational use.
V2.0.1 · retrospective evidence · not yet operational
0municipal time series modeled
0retrospective windows evaluated
0experimental weekly horizons
0%empirical IC80 coverage in the internal split
The challenge
Earlier signals, context-aware decisions
Dengue surveillance requires continuous monitoring, coordination across public-health levels and decisions under uncertainty.
Thousands of local contexts
Five macro-regions with different epidemiological patterns
Performance must be monitored by municipality, region, horizon, event frequency and outbreak conditions—not only by one national average.
Uncertainty must be explicit
Models support review; they do not replace accountable professionals
Useful intelligence combines central forecasts, uncertainty ranges, traceability and human oversight.
Official context: Brazilian Ministry of Health Arbovirus Dashboard.
How it works
From data to a probability distribution
The V2.0/V2.0.1 reference combines epidemiological and macro-environmental signals in one national Temporal Fusion Transformer.
- 01
Point-in-time data foundation
Weekly municipal records, climate signals, seasonality and vector-receptivity indicators with versioned provenance.
- 02
National shared-parameter model
One model trained across 5,570 municipal series. This does not yet mean explicit spatial propagation modeling.
- 03
Four probabilistic horizons
Central estimates and uncertainty intervals for T+1 through T+4 weeks.
- 04
Auditable delivery
Forecast, data and model versions designed for monitoring, review and rollback before any external alerting.
Current evidence
Retrospective evaluation across 267,360 windows
These are internal 2025 results. The year participated in model selection, evaluation and calibration; no untouched final test or prospective guarantee exists yet.
R² 0.9455
Global one-week retrospective coefficient of determination.
R² 0.8281
Global four-week retrospective coefficient of determination.
IC80 ≈ 79.9%
Empirical coverage observed in an internal random 50/50 recalibration split.
| Region | 1 week | 2 weeks | 3 weeks | 4 weeks | Series |
|---|---|---|---|---|---|
| Brazil | 0.9455 | 0.9157 | 0.8725 | 0.8281 | 5,570 |
| North | 0.9440 | 0.9193 | 0.8883 | 0.8551 | 450 |
| Northeast | 0.9242 | 0.8920 | 0.8609 | 0.8371 | 1,794 |
| Central-West | 0.9425 | 0.9178 | 0.8928 | 0.8729 | 467 |
| Southeast | 0.9469 | 0.9167 | 0.8729 | 0.8279 | 1,668 |
| South | 0.8926 | 0.8671 | 0.8273 | 0.7798 | 1,191 |
Retrospective internal evaluation · epidemiological source: InfoDengue · values will be re-estimated under the V2.1 point-in-time protocol.
Potential applications
B2G first, B2B as a secondary path
Public-health management is the primary market. Private-health applications remain hypotheses subject to validation, governance and workflow integration.
Health departments
Preventive planning
- Prioritize epidemiological review and field investigation
- Combine forecasts with operational indicators
- Keep final decisions with accountable managers
Epidemiological surveillance
Quantified uncertainty
- Weekly trajectories with probability ranges
- Experimental signals reviewed by epidemiologists
- Associative explanations without automatic causal claims
Public management
Auditable decisions
- Versioned data, models and forecasts
- Human review and incident runbooks
- No autonomous resource allocation
Private health · secondary
Capacity and population planning
- Hospitals and networks: capacity preparation
- Health plans: aggregated territorial intelligence
- No individual diagnosis or automated clinical recommendation
Evidence-gated roadmap
From retrospective evidence to safe operation
Later capabilities are conditional on scientific, operational and governance gates. Dates are not presented as guarantees.
V2.0 / V2.0.1
National probabilistic model and internal interval recalibration.
- Retrospective only
- Not operational
V2.1
Point-in-time reconstruction, strong baselines, rolling-origin evaluation and untouched final test.
- Scientific validity
- Updated Model Card
V2.2–V2.4
Weekly MLOps, at least 12 shadow-mode weeks, prospective pilots and locally validated alerts.
- Traceable operation
- Human oversight
V2.5 / V3
Scenario sensitivity, causal safeguards, territorial foundations and spatial models only after baselines.
- Sensitivity is not causality
- ST-GNN only with incremental evidence
V4
National platform, multi-arbovirus scope, authorized integrations and a source-grounded epidemiological assistant.
- Governance by design
- Separate validation per disease
V5
Country-specific pilots, transportability studies and conditional research on epidemiological foundation models and digital twins.
- No automatic transfer from Brazil
- Local partners and validation
Recognition
Early recognition for innovation and creativity
Awards recognize the proposal and execution; they do not replace scientific, commercial or prospective validation.
2026
Curitiba Mais Criativa
Winner · AI and Applied Creativity
2026
Founders Club Growth Challanger
Selected · powered by Notion
2026
Prêmio Brasil Criativo
National stage · Finalist
Computer program registered with Brazil's INPI · Process BR512026004146-5.
Founder
Guilherme Henrique Pereira
Data Engineer · AI, Machine Learning, Big Data and automation
Founder and developer of Dengo AI, with professional experience at Volvo Group, Votorantim and Vivo. Software Engineering student at UNINTER, DIO Expert ambassador and early-career technology mentor.
Dengo AI is a venture in formation, currently without a dedicated legal entity. Its software is registered with Brazil's INPI under process BR512026004146-5.
Next step
From retrospective evidence to prospective validation.
Review the results, current limitations and technical gates that guide Dengo AI's development.