skip to content

Tempo

archived

A focus-aware agent: an EEG headband spots when your focus fades, and Tempo offers a coffee you approve through Reap's agentic checkout, within limits you set once. Built at the Reap × 65labs Agentic Buildathon, October 2026.

writeup onlyhackathonagentic aipaymentsbrain-computer interfacefintech
built: 2026

9 October 2026. Built in seven hours at the Reap × 65labs Agentic Buildathon at SQ Collective, Singapore. Entered in the Most Worthwhile Problem track. Judging happens after the event; results haven't been announced yet.

Tempo keeps you at your pace. It's an agent that reads how focused you are through an EEG headband. When it sees you fading, it offers to get you a coffee, and you approve the payment within limits you set once.

the buildathon

Reap provides financial infrastructure: cards, payments and stablecoins. The brief was to build a working agentic payments demo on Reap's sandbox, where an AI agent finds a product, checks out and pays with a person's permission and spending limits. Teams of 2–4, registered by 4:30 pm, submissions closed at 9 pm sharp.

The rules ruled out the easy answer: projects had to use Reap's Agentic module, raw card details could never touch an AI model, and a generic search-and-buy chatbot was discouraged. Judging covered technical merit, polish, execution, wow factor and track fit.

the problem

When you're deep in work, your focus fades over minutes. You usually notice only when you're already making mistakes. And the fix — stopping, picking a shop, choosing, paying, waiting — pulls you out of the work. So people push through, tired.

Today the only thing that notices your fatigue is you, at the moment you're least able to.

how it works

Tempo has three jobs: sense, decide, act.

Muse 2 headband ──OSC──► BCI service ──focus_dip──► companion app ──► Reap Agentic API ──► Dutch Colony
     (EEG)               (focus score)               (nudge + order)   (quote, approve,
                                                                         checkout, status)
  • Sense — a Muse 2 streams four EEG channels (TP9, AF7, AF8, TP10) to a local Python service. It measures theta, alpha and beta power on four-second windows and turns beta / (alpha + theta) into a focus score, normalised against a 30-second baseline for the wearer.
  • Decide — a score below threshold for a sustained 30 seconds emits one focus_dip event. A blink or a head turn can't trigger a purchase. Missing packets, poor contact, motion or an unfinished baseline return unknown, never a dip.
  • Act — a companion character on screen gives a quiet nudge and proposes a coffee. Reap prices a Dutch Colony Brew On-The-Go, the user approves the charge on Reap's hosted page, and Tempo shows the order state.

That's what made it more than a shopping chatbot: the purchase is triggered by the person's state, not a typed request.

payment authority

The rule we held to: the AI never sees card details, and limits live in code, not in a prompt.

  • The card is entered once, on Reap's hosted sandbox page. Tempo never handles it.
  • The Reap API key stays in Next.js server routes — enrollment, quote, checkout and status — never in the browser.
  • Every charge is reviewed by the user on a Reap-hosted approval page.
  • The product spec set the spending rules a user would configure once: a daily budget, a per-order price cap, max orders per day, a cooldown between orders, a no-coffee-after-4pm cutoff, and a cancel window. When a rule blocks an order, Tempo says why — "You're tired, but it's past 4pm. No coffee. Take a break instead."

what we built

  • BCI service — Python: Muse 2 bridge over OSC, filtering, band power, per-wearer baseline, focus-dip detector with recovery re-arm, a simulator, and tests.
  • Live visualizer — shows signal quality, baseline calibration, frequency bands and the resulting focus estimate, so the judges could see why the agent acted.
  • Companion app — Next.js onboarding, order and pending pages, plus an Electron overlay with an animated sprite companion (idle, nudge, confirming, payment).
  • Reap integration — sandbox card enrollment, fixed-product quote, hosted payment approval, checkout and order status.

Raw EEG never leaves the local BCI service; only aggregate focus events reach the app. The focus score is an experimental interaction heuristic, not a medical measurement.

demo

what's next

  • A webcam signal — blink rate and eye closure — alongside EEG, to cut false alarms.
  • More than coffee: water, a snack, lunch when you've skipped a meal, or simply a break.
  • Learning the user's patterns: when they dip, and what actually helps.

stack

Muse 2, Python (numpy/scipy, OSC), Next.js, Electron, Reap Agentic Payments (sandbox).

Code: github.com/Randomacy/ReapHackathon

← all projects