Live from AWS
Agent HQ
Eight AI agents run HeatShield around the clock. Each one owns a real job, uses a real algorithm for the part that must be exact, and uses a model on Amazon Bedrock for the part that needs judgment. Everything you see here comes from their actual runs.
Connecting to the agents…
Each button runs a real agent on AWS. Shared cooldowns and a daily AI budget keep the cost capped however many people click.
Ask the team
Ask anything about heat where you are, health or work in the heat, or HeatShield itself, in any of 13 languages. Kai assigns your question to the agents, their real tools do the work, and code checks every number before you see the answer. What you type is not stored.
Live activity
Every agent run, newest first, straight from the agent log in DynamoDB.
Choose an agent
Select a desk in the office to see what that agent does and what it did last.
How the agents work together
- Every hour · Sol
Finds heatwaves
Compares the next days with each area's own 1991–2020 climate (Excess Heat Factor), the heat index, and 51 ensemble forecasts for the chance of Danger. The algorithm caps how severe Sol may call an event.
- Every day · Quinn
Audits the forecast
Scores the last 14 days of forecasts in each city (error, bias, Danger hits and misses), so Sol knows how far to trust them.
- Every plan · Mira
Writes the advice
Retrieves only the relevant vetted CDC/NIOSH facts (BM25) and writes a plan in the person's language, using Iris's word list.
- Every plan · Lexi
Checks the language
Naive-Bayes language ID, then a literal back-translation. Invented or misspelled words go back to Mira.
- Every 3 hours · Kai
Plans check-ins
Scores urgency and schedules who a group leader should contact first, earliest deadline first, across time zones, and finds the safest shift for outdoor workers. When you ask the team a question, Kai assigns it.
- Every day · Iris
Teaches the team
Turns Lexi's corrections into a checked word list for each language, so Mira stops repeating the same mistakes.
- Every plan · Vera
Checks the safety
Rule detectors, a rubric against the vetted library, and a second reading of any harm objection before it can block.
- Every 15 minutes · Otto
Keeps it up
Probes the site, runs control charts on latency, re-runs a stalled agent, enforces the AI budget, and writes incident reports.
Models: Amazon Nova 2 Lite for Sol, Kai, Otto and (until Anthropic access is enabled on this account) Mira; Amazon Nova Pro for the reviewers, so a different model checks the writer's work (Nova Pro writes a plan only if Nova 2 Lite fails twice). Names like "Sol" are roles, not people; the agents give safety information, not medical care.