Timing
Notifications describe what has already occurred
Preventive decisions must combine observed data with future signals and uncertainty.
Deep Tech · GovHealth
Artificial Intelligence
Public Health
Dengo AI
Predictive epidemiological intelligence that turns dengue data into risk signals, maps, and decision support.
5,570
municipal time series modeled
0.9455
global T+1 R² · retrospective evaluation
4
simultaneous weekly horizons
Dengue · current model
Chikungunya and Zika · V4.1 roadmap
INPI registration BR 51 2026 004146-5
Internal 2025 metrics; V2.1 is the next scientific cycle. The system does not currently operate as an official alert service.
The problem
Dengue spreads quickly. Public response needs time and context.
Surveillance data mature through revisions, while risk, seasonality, and response capacity differ across municipalities.
6.5M+
probable cases in 2024
6,321
confirmed deaths in 2024
Source: Brazilian Ministry of Health, Epidemiological Report EW 46/2024 and Information Note No. 12/2025. Data are subject to revision.
Scale
5,571 municipal realities
Volume, climate, seasonality, infrastructure, and data quality are not uniform.
Action
A forecast alone is not enough
The signal needs context, quality controls, and a clear human-review workflow.
The solution
A platform to forecast, explain, and prioritize.
Dengo connects epidemiological and climate data to a national probabilistic model and a surveillance-oriented product layer.
Public datacases, climate, territory
→
Pipelineweekly harmonization
→
National TFTT+1 to T+4 and quantiles
→
Intelligencerisk, uncertainty, context
→
Decision-makertriage and human judgment
Forecast
Four weeks
Simultaneous horizons for different planning windows.
Coverage
Municipal scale
One model jointly trained across 5,570 Brazilian time series.
Uncertainty
Seven quantiles
Experimental intervals to communicate range and risk.
Product
Dashboard and reports
Prototypes that turn forecasts into actionable information.
Technology
The engine: Temporal Fusion Transformer.
A deep-learning architecture for multivariate, multi-horizon, probabilistic time-series forecasting.
- Current 26-week encoder for each municipality.
- Epidemiological history and climate variables.
- Direct T+1, T+2, T+3, and T+4 forecasts.
- Seven quantiles for 50%, 80%, and 95% prediction intervals.
- 3,980,351 trainable parameters.
In plain language: the model learns shared patterns across Brazil without erasing differences between municipal series.
1 · Memory
Temporal context
Processes prior weeks and the recent state of each series.
2 · Attention
Relevant patterns
Weights useful temporal relationships for each horizon.
3 · Selection
Internal variables
Controls signal contribution without treating it as causal.
4 · Output
Probabilistic forecast
Provides a median and intervals for four future weeks.
Results and scientific rigor
Promising performance with documented limitations.
0.9455
global T+1 R²
267,360
windows per horizon
79.9%
80% interval after internal recalibration
| Horizon | R² | RMSE | MAE |
|---|---|---|---|
| T+1 | 0.9455 | 33.54 | 3.86 |
| T+2 | 0.9157 | 41.56 | 4.47 |
| T+3 | 0.8725 | 50.80 | 5.04 |
| T+4 | 0.8281 | 58.56 | 5.45 |
Correct interpretation: an internal retrospective evaluation against the processed 2025 target.
- The target was truncated during part of the evaluated period.
- Observed ONI introduces a potential look-ahead risk.
- No fully independent final test set is available.
- Current calibration uses temporally dependent windows.
V2.1 is designed to address these issues before operational promotion.
The next slides detail regional performance, calibration, alarms, heterogeneity, and audit findings.
Territorial results
National coverage does not imply uniform performance.
5 regions
evaluated separately across all four horizons
Brazil's South has the lowest regional R² at every horizon; V2.1 must investigate regime, data, and generalization.
R² against the processed target · internal retrospective evaluation from 2025.
| Region | T+1 | T+2 | T+3 | T+4 | 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 |
T+1
RMSE 33.54
MAE 3.86
T+2
RMSE 41.56
MAE 4.47
T+3
RMSE 50.80
MAE 5.04
T+4
RMSE 58.56
MAE 5.45
Reliability and audit
Point accuracy, uncertainty, and methodological risk.
Probabilistic coverage
| Horizon | Raw 80% | V2.0.1 80% | V2.0.1 95% |
|---|---|---|---|
| T+1 | 27.50% | 79.94% | 94.92% |
| T+2 | 26.27% | 79.88% | 94.92% |
| T+3 | 25.43% | 79.79% | 95.02% |
| T+4 | 24.94% | 79.90% | 94.91% |
Experimental alarm · national P90 threshold
| Horizon | Precision | Recall | F1 |
|---|---|---|---|
| T+1 | 0.9395 | 0.7609 | 0.8408 |
| T+2 | 0.9360 | 0.7103 | 0.8077 |
| T+3 | 0.9301 | 0.6625 | 0.7738 |
| T+4 | 0.9221 | 0.6256 | 0.7455 |
3,920
low volume
1,068
R² ≥ 0.50
336
moderate
246
weak
A1 · Critical
Ground truth modified in 5,121 rows from 2025.
A2 · Critical
Revisable ONI treated as known future data.
A3 · Critical
No independent final test set.
A4 · High
Random split with temporal dependence.
Recalibration corrects coverage on the internal split but provides no prospective guarantee. The alarm is not an official outbreak definition.
Product and application
From the curve to decision context.
The prototype brings forecasts, intervals, quality, and human review into one experience.
Health departments
Municipality triage, investigation planning, and risk interpretation.
Surveillance
Signal, local threshold, observed data, and uncertainty in one workflow.
Regional management
Comparable views to prioritize support across territories.
Dengo AI · Municipal intelligenceOrigin: EW 18
T+1predicted median
80% PIprediction range
Reviewexperimental status
Signal for reviewThe decision-maker confirms freshness, trend, and context before taking action.
Operational benefits still require prospective measurement; the current dashboard is a decision-support prototype.
Market and business model
B2G as the entry point. B2B as expansion.
5,571
Brazilian municipalities in the 2025 territorial division
27
federative units and regional management levels
These figures represent potential administrative units, not contracted customers.
Current focus: municipal, regional, and state public-health surveillance. A shared architecture can expand coverage without maintaining a separate model for every customer.
Core · B2G
Public management
Institutional licensing, deployment, monitoring, and surveillance support.
Expansion · B2B
Hospitals and networks
Capacity planning after integration and validation with admissions, demand, and catchment areas.
Expansion · B2B
Health insurers
Regional intelligence and network planning with aggregated data, defined purpose, and governance.
Long term
Industry and APIs
Analytics for pharmaceutical companies and laboratories, plus APIs, only after specific validation.
Potential revenue: B2G licensing, B2B analytics subscriptions, and scope-based APIs.
The current model forecasts municipal cases—not admissions, claims, or pharmaceutical demand. Each vertical requires new data and validation.
Assets and differentiation
An advantage built across data, model, and product.
Dengo does not replace the official ecosystem. Its thesis is to integrate technical and operational capabilities into an auditable platform.
Data
National pipeline
Harmonized weekly series, climate, and territorial context.
Model
Probabilistic forecast
National TFT, four horizons, and seven quantiles.
Product
Decision workflow
Dashboard, experimental alerts, reports, and human records.
Intellectual property
Registered software
Computer program registered with Brazil's INPI.
| Dimension | Current state | Next strengthening step |
|---|---|---|
| Scale | 5,570 modeled series | Governed weekly operation |
| Evidence | Internal retrospective evaluation | Independent and prospective testing |
| Product | Functional prototype | Usability and human factors |
| Business | Early recognition | Commercial validation and deployment |
InfoDengue, Mosqlimate, SINAN, and public dashboards are sources, references, and potential integrations.
Founder and recognition
Guilherme Henrique Pereira
Data Engineer · Founder and developer
Hands-on expertise in AI, machine learning, big data, automation, cloud computing, and scalable systems.
PythonML & Big DataCloudBackendDIO Expert
Software Engineering undergraduate at UNINTER, Brazil.
Volvo Group
Python, generative AI, and document-analysis automation.
Votorantim
SAP–ServiceNow integration and operational efficiency.
Vivo
Java Spring Boot, databases, and backend performance.
Winner
Curitiba Mais Criativa
AI and Applied Creativity · 2026
Finalist
Prêmio Brasil Criativo
National stage · 2026
Selected
Growth Challanger
Founders Club
Roadmap · Foundation and operations
From internal evidence to safe prospective use.
✓ V2.0 / V2.0.1
Retrospective foundation
National TFT, T+1 to T+4, seven quantiles, local diagnostics, and internal recalibration.
Now · V2.1
Scientific validity
Point-in-time data, raw target, untouched test set, baselines, ablations, and subgroup metrics.
V2.2
Operations and MLOps
Weekly ingestion, versioning, quality gates, observability, rollback, and shadow mode.
V2.3
Prospective pilot
Frozen forecasts, safety, usability, human factors, and operational metrics.
V2.4
Local alerts
Incidence, endemic channel, territorial thresholds, investigation, and structured feedback.
V2.5
Scenarios
Climate sensitivity; interventions and ROI only with appropriate causal identification.
Central gate: each version advances only after meeting predefined scientific, operational, and human criteria.
Roadmap · Platform and expansion
A broad future, built through verifiable layers.
V3 · Territorial intelligence
Understand relationships between cities
- V3.1: spatial and mobility foundation.
- V3.2: graphs/ST-GNN after baselines.
- V3.3: Dengo Visus in a supervised track.
V4 · National platform
Expand diseases and integrations
- V4.1: dengue, chikungunya, and Zika.
- V4.2: B2G and B2B integrations with hospitals, networks, and insurers.
- V4.3: auditable epidemiological assistant.
- V4.4: recommendations under protocols.
V5 · Global intelligence
Expand with local validation
- V5.1–V5.3: Latin America, Africa, and Asia.
- V5.4: epidemiological foundation model.
- V5.5: epidemiological digital twin.
V3 depends on demonstrated incremental spatial gain.
V4 requires disease-specific governance, data contracts, LGPD compliance, and validation by vertical.
V5 is a long-term vision, not a contracted deliverable.
Scale and impact
From Brazil to an infrastructure for epidemiological intelligence.
Brazil
Municipal dengue
Validate operations, utility, safety, and sustainability within a complex surveillance system.
Multi-vertical
B2G and B2B
Reuse infrastructure across new journeys without assuming one outcome serves every market.
Multi-country
Adaptable framework
Expand through data contracts, local context, and territory-specific evaluation.
SDG 3
health and well-being · impact to be measured
SDG 10
territorial equity · subgroup performance
SDG 13
climate and health · association without automatic causal inference
Reductions in cases, hospitalizations, costs, or mortality remain impact hypotheses until supported by an appropriate prospective evaluation.
Vision
Forecast better. Decide with context. Evolve through evidence.
Dengo AI combines nationwide data engineering, probabilistic artificial intelligence, and a validation-gated roadmap to build a new layer of epidemiological surveillance.
5,570
municipal series modeled
V2.1
next scientific gate
V5
conditional global vision
Dengo AI · Guilherme Henrique Pereira · computer program registered with Brazil's INPI.