Tempo arrows or click to advance · N for speaker notes

For Gabe · from Jeff

Tempo.

Your Sana bet, extrapolated: from one person building fast with AI to a team building fast together.

Jeff Williams1

Where this starts

Your experiment worked. That sets up the next question.

  • The bet: one person with AI can build an EHR from scratch. You made it, you doubted it, and you answered it.
  • The next question arrives with scale: as more of the team joins Sana, speed stops being about one builder and starts being about not waiting on each other.
  • This proposal: the same bet, extrapolated to a team, run the way Sana itself started: as an experiment, in steps, measured.
  • And it aims at Friday's numbers: the market now pays for AI that's real in the product and efficiency you can show.
Building on the bet2

This already happened, and you merged it

One Sana feature shipped while nobody stayed up.

4 days
Thursday-night brief to Sunday merge, measured
$45
compute for the shipped build
~2.5 h
total human time in the cycle

Built and tested overnight Thursday; iterated and reviewed over the weekend; merged Sunday. Your review found real problems, every one was folded in before merge, and each one became a learning we fed back into the process for every run after.

The check-in stops cycle3

The problem worth solving

Building got fast. The waiting didn't.

Work waits

When teams measure where the time really goes, active work is under 15% of it. The rest is waiting. AI sped up the work and left the waiting untouched.

Decisions wait

One unanswered question can hold three branches. The decision, not the code, becomes the bottleneck.

Overlaps happen

Rarer than it sounds, but expensive when it lands: a shared contract, a generated file. Worth cheap insurance that catches it early.

The waiting is the opportunity4

The shape, and the name

Tempo: four parts, one pace.

the next night Nightshift the build loop: research, build, verify with fresh eyes, package a demo and a PR. Overnight, unattended. Daybreak the morning side: the operator reviews the night's work, makes the calls, merges or redirects. Human. Dispatch the coordination layer: one ranked list, set weekly by the group; decisions routed to their owner, answered same-day Pulse the scoreboard: every claim measured

Loops inside loops: nothing assumes the first pull request is perfect. The system assumes the opposite, and every review makes the next run better.

Nightshift · Daybreak · Dispatch · Pulse5

What keeps it safe, at this stage

Nothing merges itself.

  • A person approves every merge. Same bar as today. In the parity phase, with no customers on the platform, this is the bar that matters.
  • Fresh eyes verify. A separate session that didn't write the code checks it against the brief, then re-checks after fixes, before any person spends time on it.
  • Hard lines wired in from day one. Money, patient data, tenant separation, access control: stop-and-ask reflexes now, long before real clinics and real data arrive.
  • Autonomy is measured, never assumed. Interventions per run are counted; the leash lengthens only as evidence earns it.
Structure over discipline6

The part you said out loud

"We are humans, and our workflows should be mindful of that."

The team's most-voted retro item, and the design brief for this. Their top concerns each shaped a piece of it:

"Losing connection to the code"

Depth is a personal choice: read every line, replay the night, or trust the checks. Watching closely at first is the intended on-ramp.

"Pressure to fit one style"

The checkpoints are fixed; everything between them stays personal: your tools, your style. Improvements you file become everyone's.

"Overlap of efforts on Sana"

Their own retro worry, answered mechanically: branches declare intent, the check compares, shared pieces land first.

Their retro is the design brief7

The answer to "no sustained gain yet"

Run it on a scoreboard, in the open.

64%
shipped features clearing review first pass
60 of 321
customer asks addressed by shipped work
21%
merges later needing a fix, counted in full
$52
per feature, only if it survives 90 days

Example numbers, marked as such everywhere; the 17 → 3 intervention curve, from the first cycle to the one that merged, is the real one. The process is measured too: interventions per run, review rounds, improvements adopted.

Pulse: outcomes first, example vs measured always marked8

The experiment, continued

One small crew. One quarter. Measured in the open.

  • The first pass, concretely: the next Sana Mondays group; two or three engineers who opt in, each running the overnight loop on gap-list features; I run the process and the measurement around them.
  • Success stated up front: the gap list moves, quality holds, cost per feature stays in the tens of dollars.
  • Kill switch stated up front: if quality or cost breaks and stays broken, we stop and say so.
  • Division of labor: you and the leads keep every review and every call you make today. I carry the process, the measurement, and the reporting.
Steps, signals, then scale9

If you want to go deeper

The path, in the order I'd read it.

  1. Tempo: the full design · how a team runs the loop without waiting on each other. A ten-minute read, with the mechanics behind it when you want them.
  2. Pulse · the scoreboard, live on example data. Every number clicks through to how it's computed, where it comes from, and how it could lie.
  3. Usage measurement inside Sana · the proposal for instrumenting the product itself the same way, from the first screen.
  4. The nuts and bolts · the Tempo repos in our GitHub: the loop process, the coordination scripts, the team vault. Clone them and look.

Link and passphrase are in my message. No meetings required to go any level deep.

Jeff Williams10