Anna

The core ideaHow she's actually smart

Anna isn't smart because she has access to a good language model. She's smart because she understands one specific person, in increasing depth, over time.

Every signal — what she hears, where she is, how her body's doing, who's around, what tempo fits the moment — feeds one map, not seven separate features that don't talk to each other. The map learns. The longer it runs, the less it has to ask.

How it worksThe calibration model

Week 1Anna hears an activity and guesses. "Cooking?" Roughly half the time, she's right — this is the honest starting point, not a demo number.
Week 2Audio connects to place. The same sounds mean something different in a kitchen than in an office.
Week 3Heart rate joins in. Cooking at 7am calm reads differently than cooking at 3pm stressed.
Week 4Social and route patterns join. A Tuesday afternoon text plus a familiar route stops being a coincidence and starts being a prediction.
Month two and onThe map is dense enough that she starts speaking before being asked. That's the actual target — not a chattier assistant, a quieter one that's usually already right.

One map, not seven featuresWhat's actually connected

AUDIO      → what are you doing?
GPS        → where are you?
HEALTH     → how are you holding up?
SOCIAL     → who's with you?
MUSIC      → what tempo fits this?
ROUTES     → how long, realistically?
PREDICTIVE → what comes next?

All connected. All learning. All reinforcing each other.

A concrete exampleDriving to a friend's house

Early on

Anna hears an engine. Guesses "driving," about half-confident. Plays a generic driving playlist — correct in shape, blind to everything else.

Once the map has depth

She's not just hearing an engine — she's cross-referencing the drive signature against the route, the day, an incoming text, and a slightly elevated heart rate that reads as focus, not stress. Orchestral, no lyrics. Timed to the turns. A passenger gets in, volume ducks automatically. Nothing was asked. It just already knew, because it was the same route that worked last time.

The calibration is the intelligence. Nothing here is a separate "smart feature" — it's one model getting less wrong.

Why one map beats fragmentsThe actual comparison

The fragmented way — ask for driving music, get a generic playlist that has no idea what today actually is:

"play driving music"
→ generic playlist
→ doesn't know the stress level
→ doesn't know who's about to get in the car
→ doesn't adjust for a passenger

The unified way — nothing was asked at all:

orchestral, no lyrics   (focus drive, not distraction)
passenger-aware         (volume drops when someone gets in)
the same music          that already worked, last Tuesday

Where it standsThe roadmap

Phase 1 — Voice foundation shipped

Runs on a Series 11 watch and an iPhone. Claude responds, music plays, the confirm/correct loop is live and actually learning.

Phase 2 — Calibration

Audio and place fuse together. Health context joins. Accuracy compounds instead of guessing cold each time.

Phase 3 — Social awareness

Passenger detection, learned route patterns, music that conducts itself to the actual drive instead of just playing behind it.

Phase 4 — Proactive the actual goal

The map is dense enough to speak first. Not a louder assistant — the point where asking becomes the exception.

Beyond the watchReal hardware, in progress

The watch and phone are the software proof. What's actually being built next is physical.

Palm Display Concept locked

Flip the watch face-down, slide the band up the forearm — a deliberate two-part gesture that can't happen by accident. At full stretch, a MEMS laser projector on the case back throws a real display onto your own palm, with finger-position tracking turning each finger into an addressable button.

Normal mode: a plain watch, face up, nothing unusual visible to anyone nearby. Private mode is one gesture away.

Triangle Puck In development

A paired pair of small camera-and-laser devices, each reporting to the Mac Mini as the brain — the same pattern Anna's tools server already uses for the watch. One puck can point at something it can see; two that know their own position in a room can triangulate, holding a level line steady or resolving a real 3D point instead of one puck guessing alone.

Find things — point a camera and a laser at a named object across the room.
Hold a reference — a level line held across a wall for tiling or hanging a frame, instead of a laser level and a guess.
Point at a task — hold a dot on a part or an instruction point while both hands are busy elsewhere.

A companion for people who live alone Early build order

The furthest-out thread: real company for someone who's isolated — personality and dark humor, not a chirpy assistant — that finds things, notices if someone's gone quiet too long, and is never the single point of failure for a real emergency. Built value-first: a good box on the counter before anything ever tries to fly.

Why ClaudeShe sees the work, not just the words

Other assistants see whatever you type. Anna's model has access to the actual body of research and music behind begump — the framework, the songs, the years of pattern — sitting in memory, not buried in generic training data.

That's the real difference between a model answering a question and a model that already knows the person asking it.

Privacy & securityLocal by default, real encryption, no quiet exceptions

Anna doesn't upload memories or health data anywhere, ever, by default — and nothing about that changes silently after an update.

Encrypted at restAES‑GCM on every memory and health file, key held in the Secure Enclave, unlockable only on that device.

Explicit opt‑inCloud reasoning, the Mac's LAN tools, and the site relay are each off until switched on individually — nothing bundles.

Logged locallyEvery network request Anna makes is written to a local log on-device. Never transmitted anywhere.

Audited machine-sideA standing local monitor watches for exactly the kind of accidental exposure — a port left open to the world — that got fixed once and is checked for ever since.

His API key. His device. His control.