What they did
During four weeks of parental leave, Navdeep Jain, who leads a guardrails team for a financial firm's AI platform, built Baby Watchdog by hand against Anthropic's API with Claude Haiku and no agent framework. He and his wife post in a private #baby-log Slack channel in plain words like a feed amount and time or a wet diaper. The bot turns each message into a structured event in DynamoDB, checks it against a rolling baseline, posts an alert when something is overdue (for example no feed in five hours), and sends a daily 9am digest and weekly summaries. Trend questions go to a tool loop where Claude picks which of three lookups to call, up to five rounds. A two-minute emoji approval guards the one action that can undo data, with silence counting as no, plus an audit log, a kill switch and Bedrock Guardrails screening each message.
How it went
The household used it daily; the post shows a full real day of logs, an overnight gap alert and a database-backed trend answer. It first ran always-on in Bedrock AgentCore at about $1.25 a day, then he rebuilt the Slack side as three Lambda functions and it dropped to about $0.02 a day. Things broke along the way: a silent two-hour outage when AWS idled the instance, daylight-saving time errors when Claude did the time math, "4 wet diapers" collapsing into one event, and messages lost during the migration, which he recovered by replaying Slack history.
Worth knowing
His main lesson: keep arithmetic out of the model. Claude now reports the local time and amounts it heard, and plain code does time zone and unit conversion. A recovery feature built for one architecture did not carry over to the new one, so check your safety nets after any rebuild.