Blog Engineering · Data

I tried to predict disease outbreaks with AI. I failed twice. Here's what I built instead.

2026-05-0110 min readAlex StepikinAIEpidemiologyData EngineeringNetlify FunctionsOpen Source

The original idea

In early 2025, I had a hypothesis: climate signals and media noise could predict infectious disease outbreaks before official surveillance bodies reported them. If true, even a 2-week advance warning would be enormously valuable for travel risk assessment, supply chains, and public health response.

I pre-registered three backtests before touching the data. Pre-registration matters: it's the only way to distinguish a real finding from one you shaped after seeing the results. The commitment was simple — if the models don't work, I publish the failures.

Experiment 1: Dengue + climate signals. Failed.

Hypothesis: ENSO anomalies, sea surface temperature, and precipitation data could forecast dengue outbreak timing at country level, 4–8 weeks ahead.

Result: the model performed worse than the climatological mean. Climate variables had no predictive skill for dengue at country-level resolution. The signal existed in the literature — but it existed at regional granularity, not at the political-boundary level needed for actionable alerts. Country borders don't respect ecological zones.

The model was not wrong. The hypothesis was wrong. That distinction matters.

Experiment 2: Media volume → outbreak proxy. Failed.

Hypothesis: spike in news volume about a disease in a region predicts an upcoming WHO alert. Used ProMED, GDACS, and Reuters RSS feeds as the media layer.

Result: media volume correlated with official reports — but lagged them, not led them. The news was reporting outbreaks already recorded by surveillance bodies. The signal I needed didn't exist in the media layer.

The honest pivot

Both experiments killed the predictive claim. What the data could honestly support: transparent, source-attributed aggregation with an auditable risk methodology. No black boxes. No proprietary scores. Every number traces back to a primary source.

This became Vigilo.

Vigilo global threat radar — 198-country disease surveillance dashboard with live source tracking

What Vigilo actually is

45 sources — WHO, CDC, ECDC, ProMED, PAHO, PHAC, GDACS, ReliefWeb, and 36 others — parsed on a 12-hour cadence via GitHub Actions. An exposure graph with 108 nodes and 458 edges models how outbreaks propagate between countries and sectors.

198-country coverage on a 3D globe with a heatmap toggle. Severity scale 0–5, live event markers, drill-down to primary source reports. The methodology for every risk score is published in the docs.

26 Netlify Functions as microservices handle ingestion, normalization, alerting, and the API. Alert delivery via Resend transactional email, @vigilocc_bot on Telegram, and VAPID web push.

Why transparency is the product

The market for pandemic intelligence tools is dominated by products with black-box risk scores and no disclosed methodology. Bluedot, HealthMap, and others produce scores you can't audit. That's a problem in a domain where the cost of false confidence is measured in lives.

Vigilo's differentiator isn't the algorithm — it's the audit trail. Every alert shows which of the 45 sources triggered it, which data fields matched the alert criteria, and a permalink to the primary report. You can disagree with Vigilo's assessment. You have everything you need to do so.

The infrastructure decision

Running 45 parsers on a schedule, normalizing heterogeneous data formats, and serving a real-time map to concurrent users — this is not a simple architecture. The decision to use Netlify Functions as microservices (rather than a monolithic server) was a cost and reliability choice: functions scale to zero when idle, and each parser failure is isolated.

The 12-hour cadence is intentional. Most surveillance bodies update daily. Parsing every hour would add cost and noise without adding signal.

What I published

Both failed experiments are documented. The methodology is public. The source list is auditable. The code is on GitHub.

Vigilo is live at vigilo.cc. The pre-registered backtests are published in the case study at stepikin.com/cases/vigilo.

Building an AI product and need design + frontend shipped fast? That's exactly what Stepikin Studio does — concept to live product in 2–4 weeks.

More from the studio

Create bold.
Deliver better.

See our workGet in touch