Firstday Go

Your trainer said it once. Bee heard it. FirstDay Go turns their exact words into practice, so new hires with ADHD never have to ask twice.

Comment

Kuto, 26, has ADHD and starts at a café on Tuesday. At 9:10 his trainer, Rosa, walks him through twenty small rules while the espresso machine screams. At 2 pm a customer hands him a cracked mug, and he can't remember whether to check the receipt first or which bin it goes in. He asks Rosa again, for the third time that day. The mug isn't the cost. Looking like he can't be trusted with the job in week one is. (Kuto is an illustration.)

FirstDay Go turns what your trainer actually said, recorded by Bee on your Apple Watch, into calm practice for your next shift. Every step keeps the trainer's exact words, a "maybe" never becomes a rule, and when a rule changes, everyone who practised the old way gets a Change Drill.

Real Bee data, filmed live: Bee records the training on an Apple Watch, Bee's cloud processes it, FirstDay's Bee service on AWS Lambda brings it in, the same conversation appears in FirstDay Go, and Amazon Bedrock finds the work steps with Rosa's exact words.

Real data, no mock-ups. Filmed on October 18, 2026: Apple Watch → Bee → FirstDay Go → Amazon Bedrock.

Inspiration

A first day is mostly talk. A trainer explains twenty small procedures in an hour, and a new hire is expected to remember them. That's hardest for the people we built this for:

  • 15.5 million U.S. adults have a current ADHD diagnosis (CDC, 2024).
  • Adults with ADHD do worse on verbal, not visual, long-term memory (Skodzik et al., 2017). Spoken training is the weak channel.
  • Food-service workers quit at 4.2% a month, about twice the all-jobs rate, so new first days come around constantly (BLS JOLTS, 2025).

Bee already hears that conversation. Nobody had turned it into something you can practise. We also wanted the employee to own it: you use your own Bee, with your trainer's OK, and no employer has to buy anything.

What it does

The app has one job, getting you ready for your next shift, and it shows one next thing at a time.

From a Bee conversation to work steps you trust

  • It arrives on its own. With your Bee connected, FirstDay Go's Today screen shows "Just recorded by Bee" as soon as Bee finishes processing a conversation. This comes live from bee stream. Nothing is read until you tap Yes, read it.
  • Amazon Bedrock finds the work steps. Each step sits next to the trainer's exact quote, and you confirm it: Yes, that's what was said, Fix it or Not a rule.
  • Vague talk becomes a question, never a rule. "Maybe we'll start giving refunds on Sundays too" goes to Questions, to ask the trainer.
  • Bee's own data comes along. Bee's action items become to-dos and Bee's facts become memories. Both sync both ways with Bee.

Practice built for ADHD

  • Short role-plays from the job. A line-drawn customer describes a situation, and the feedback always quotes the trainer.
  • Nothing to lose. There are no scores and no streaks.
  • Steps come back on a schedule. Spaced review repeats each step after 1, 3, 7 and 14 days.
  • A practice reminder before your shift.

Change Drill, for the whole team

  1. A trainer shares a training with a 6-character code.
  2. The trainer lets each person in, and sees the verified email of each person who asks.
  3. When the trainer sends a change, everyone who practised the old way gets a Change Drill. It shows the old rule next to the new one, and the old answer becomes the trap choice, so the habit actually changes.

My record, for reviews and promotions

  • Praise is kept word for word. Praise and thanks in a Bee conversation ("Great job handling that refund this morning") are suggested for your record, with the date.
  • You confirm each one. You keep each quote and say who said it, because Bee labels every speaker "Unknown".
  • Learning counts too. What you learn is added as you go ("Learned all 3 steps of Juniper Café").
  • Copy for my review gives you dated proof, in other people's words.
  • It's yours. The record stays on your phone and is shared only if you copy it.

Quiz me anywhere

Your confirmed steps power our own MCP server (spec 2025-11-25, Streamable HTTP, 6 tools and a shift warm-up prompt), the firstday-coach Agent Skill and a voice screen. Any AI assistant that speaks MCP can quiz you in your trainer's words.

How we built it

Animated architecture: Apple Watch, Bee cloud, Bee service on AWS Lambda or a computer helper, FirstDay Go, Amazon Bedrock, Amazon Cognito, the coach service with DynamoDB and the MCP server, S3 and CloudFront, and the trainees, voice screen and AI assistants it reaches.

Bee, through its official CLI, called at runtime

Bee CLI What FirstDay Go does with it
bee stream --json Live "Just recorded by Bee" on Today, plus a notification
bee conversations list / get --json Reads a training, only after you OK it
bee facts … (list, confirm, update, create, delete) Bee facts become memories, edited both ways
bee todos … (list, suggestions, create, complete) Two-way to-dos and suggestions
bee daily list, bee insights list Bee's daily summary and insights on Today

There are two ways in:

  • Connect Bee, no computer. FirstDay's Bee service runs Bee's CLI inside AWS Lambda. Your Bee sign-in is sealed with AES-256-GCM using a key that only your phone holds, and Disconnect deletes it. This is a preview, open only to the owner's account until Bee confirms server-side use. It's the route filmed in the demo.
  • A small helper on your own computer (Settings → Brain in the app). It runs the signed-in CLI and pairs with the phone through a 6-digit code. Anyone with a Bee can use this route today.

Amazon Bedrock (Nova Pro)

  • One reliable call. A Lambda Function URL calls the Converse API with a forced tool call against a JSON schema, at temperature 0. That returns work steps with their exact quotes, questions for the trainer, to-dos and memories in a single call.
  • Cheap and fast. About 1,100 input and 230 output tokens, about 3 seconds and about $0.0016 per conversation.
  • Locked down. The IAM role can invoke only Amazon Nova models, and transcripts are never logged.

The AWS stack (CloudFormation, us-east-1, pay per use)

Piece AWS
Reading conversations Lambda → Amazon Bedrock (Nova Pro)
Accounts Amazon Cognito; every Lambda checks the access token
Team trainings and the MCP server Lambda + DynamoDB (on demand)
Web app and voice screen S3 + CloudFront
Connect Bee without a computer Lambda running Bee's CLI

The app

  • Built with: Expo SDK 57, React Native and TypeScript.
  • Runs on: iPhone (the web app in Safari, or Expo Go), an Android APK, and the web through S3 and CloudFront.
  • Tested: the app's core logic is pure TypeScript with 75 tests. 39 integration tests cover the helper, the Bee service, the coach service and the MCP server, using a stand-in Bee CLI. All 114 run in CI.

How accurate is it?

A wrong step is worse than a missing one, because you'd practise the wrong thing. So we measured both, on 10 held-out workplaces written after tuning, and ran Bedrock three times because its answers vary:

Reader Precision Recall Vague lines made into rules Quotes word for word
Amazon Bedrock, Nova Pro, 3 runs pooled 94.7% 87.7% (81.5–92.6% per run) 0 of 12 100%
On-phone rules (offline fallback) 94.4% 63.0% 1 of 4 100%

Every kind of mistake it makes is published, with real outputs, in Accuracy and failure cases. One example: it kept only half of a two-part rule. That's why every step is shown with its quote and has to be confirmed. The real Bee recording gave 3 of 3 steps. It's one recording, so it shows the path works on real data; it isn't an accuracy figure.

Challenges we ran into

  • Bee's real data didn't match our assumptions, twice.
    • The timing fields weren't epoch intervals, so our first adapter rejected real conversations. We rebuilt it on a real payload.
    • conversations get wraps its reply in { conversation }. Our new importer missed that and rejected every real conversation, until a test built on the real reply shape caught it.
    • Both cases are now pinned by regression tests.
  • Running Bee's CLI in Lambda.
    • The npm package is about 420 MB, so we ship only the 37 MB Linux program.
    • Bee's sign-in lives as files in ~/.bee. The Lambda restores it for each request from a copy sealed with a key held on the user's phone.
  • Bee labels every speaker "Unknown". So you name the trainer when you save a training, and confirm who said each piece of praise on your record.
  • Bedrock varies run to run. On the same 10 transcripts, recall ranged from 81.5% to 92.6%. So we report all three runs and the pooled figure, not the best run.
  • We cut what didn't work.
    • Alarms on Bee to-dos didn't buzz on the Watch in our tests, so practice reminders are phone notifications.
    • "Live cues during a shift" were cut: Bee processes a conversation after it ends, so a cue would arrive too late.

The Bee and AWS problems are written up, with what we'd ask the tool makers for, in our 12-entry friction log.

Accomplishments that we're proud of

  • The full loop runs on real data. It goes from an Apple Watch to Bee to FirstDay Go to Bedrock, filmed live with real audio.
  • Judges can run real Bee data without a Bee. Try a real Bee recording replays Bee's own transcripts of the filmed recordings, and Bedrock reads them live in your own account.
  • Measured on workplaces it never saw, with the failures published, not hidden.
  • Nothing happens without a person's yes: reading a conversation, making a rule, letting someone into a team (unless the trainer opens it to anyone with the code), sending a change and keeping praise.

What we learned

  • Build tests from real payloads. Both importer bugs came from guessing Bee's data shape; tests built on the real shape caught them.
  • Exact quotes are the trust feature. With the trainer's own words next to every step, you can check a step in seconds instead of taking the AI's word for it.
  • For ADHD, less on screen beats more features. One next thing, no scores, and nothing to lose.

What early users said

  • "Important details, blockers, and action items can easily get lost in fast-moving spoken conversations. It helps reinforce key information by turning explanations and conversations into a more structured learning experience." — Anslem

  • "The latest personal achievement documentation feature was a nice addition. It helps me keep track of what I accomplished and makes it easier to document my contributions over time." — Huy

What's next

  • "Connect Bee, no computer" for everyone, once Bee confirms server-side use.
  • Pilots with job coaches and supported-employment programmes, who already train people one-to-one.
  • Practice by voice on an Echo, using the MCP server we already run.
  • Store releases: TestFlight and the Play Store.

Built With

Share this project:

Updates

Submission history