SENTINEL · FLU WATCH EARLY-WARNING Ingest· Score· Explain· Approve· Log
← → arrow keys to move  ·  press A to autoplay

Discover · Build · Simulate · Deploy · Operate

Remember past seasons.
Detect this season's anomaly.

Flu Watch turns real WHO FluNet history into memory: CockroachDB holds multi-year season baselines, Bedrock explains anomalies in plain English, and a human approves every alert before it reaches the field.

295
real season_baseline rows
2
CockroachDB tools used
1+
AWS service (Lambda + Bedrock)
$8,750
prize pool · Team Cloud Catalyst
CockroachDBManaged MCP Server Distributed Vector IndexingAWS Lambda AWS BedrockWHO FluNetHuman-in-the-loop

The problem

Flu response is reactive.

Health teams act after cases spike. Manually scanning surveillance data delays response, and a missed signal costs lives while a false alarm burns scarce trust and capacity.

Cases spike→ Notice late→ Scramble crews→ Outbreak grows

Flu Watch flips it: score first against real history, explain the anomaly in plain English, and put a named human in the loop before anything ships.

Data foundation · real, not synthetic

Built on real WHO FluNet history.

Every score traces back to actual WHO FluNet surveillance data — loaded into CockroachDB as a multi-year season baseline, not a demo dataset.

Flu Watch architecture

System architecture: WHO FluNet ingest → CockroachDB scoring → Bedrock explanation → human approval → audit log.

SQL Shell row count

CockroachDB SQL Shell — 295 real rows in season_baseline.

CockroachDB tools · confirmed live

Two tools, meaningfully integrated.

Hackathon rules require at least two CockroachDB tools. Here's the proof: the Managed MCP Server connection, and a distributed vector index rebuilt at 1024 dimensions to match Bedrock Titan V2 embeddings.

Managed MCP Server config
Vector index on season_baseline

Left: Managed MCP Server connection. Right: SHOW INDEXES confirming the rebuilt VECTOR(1024) index on season_baseline.

Sentinel AI · AWS Bedrock

Turned into a brief a person can act on.

Why is this region flagged? Bedrock explains it in plain English, grounded in the actual scored numbers — no hallucinated advice.

Real anomaly result
"Testville reports 120 cases against a 4-year average of 45 — a 2.7x jump with no rainfall change, consistent with an outbreak signal rather than seasonal noise." — alert_level: ALERT

Officer of the Day · the differentiator

No alert ships without a human.

Every flagged anomaly lands here, in front of a named reviewer, with the full explanation attached. Nothing reaches the field on the model's word alone.

Approval UI

Approve / Reject gate — Testville week 31 (WATCH) and week 32 (ALERT), both pending human review.

Audit trail · the decision, written back

Every decision, logged and accountable.

Once approved, the decision writes straight back to CockroachDB — who reviewed it, when, and what they decided. A real audit trail, not UI theater.

Decision log

Open source

Public repo, real history.

MIT-licensed, public, with real commit history behind it.

GitHub repo

From signal to accountable action

WHO data → CockroachDB score → Bedrock brief → human approves → audit log.

WHO FluNet data
→
CockroachDB scoring
→
Bedrock explanation
→
Officer approves
→
Audit log write-back

"No warning reaches the field without a human decision — logged, timestamped, accountable."