System architecture

Inside the quiet engine that stands in.

No magical inference. No ambient surveillance. Just structured context that people publish on purpose — and a small set of rules about what to do with it.

01
Declared input
02
Scoped reps
03
Hard constraints
01The bargain

Declared, not observed

Most productivity AI tries to reverse-engineer context by watching everything: keystrokes, DMs, screen time, mouse movement. That creates noise, anxiety, and a steady background rate of hallucination.

StandIn refuses that bargain. It only knows what someone deliberately publishes in a wrap, plus what a linked system of record like Jira or GitHub already contains.

A

Observer model

Ingests everything. Guesses intent. High noise, low trust.

B

StandIn model

Pulls work from connected tools into a draft. The author reviews and approves before it publishes. Nothing enters the record without a signature. Zero guessing.

02Where it lives

Infrastructure, not a destination

A standalone app is just another silo to check. StandIn is built as a layer that sits underneath the tools your team already lives in.

Interface

Where your team already is

No separate app, no second inbox. StandIn lives in Slack, in your code host, in your ticket tracker. The surfaces you already check are the surfaces it answers on.

Canon

One canonical record

Drafts pull from Jira, Linear, GitHub, and your calendars. The published wrap becomes the source of truth — not a duplicate to maintain, but the thing teammates get quoted back to them when they ask.

03Three representatives

Three representatives, three jobs

One model for everything is how you get confident nonsense. StandIn runs scoped representatives, each answering only for what it actually knows.

i

Personal representative

Built from one person's published wraps and the tickets they own. Answers questions the way they would — or stays quiet.

Scope — one user
Auth — personal
Function — handoff, status, handover
ii

Project representative

Scoped to a squad, service, or initiative. Pulls from every wrap on that surface and cites each contributor individually.

Scope — project or squad
Auth — org-shared
Function — state of the work
iii

Routing representative

Sits above the others. Decides which representative should take a question, or points to a human when no rep has coverage.

Scope — organization
Auth — org-shared
Function — who, not what
04 · Refusals

The constraints are the product

Most AI adoption stalls on privacy. We didn't fix that with a policy page. We fixed it by making the tool incapable of doing the things people worry about.

Context has a half-life

Wraps decay. A note from three weeks ago carries less weight than one from this morning, so stale context doesn't masquerade as current.

Your words stay yours

No wrap is sent to train a foundation model. Retrieval is in-context, per question. When the conversation ends, nothing persists in anyone's weights.

05Second-order effects

What changes when presence isn't the point

Decouple context from the person who created it and some things get quietly better.

a.
Time zones stop mattering.

A question at 2am in Berlin gets the same answer as a question at 3pm in San Francisco. It's the wrap answering, not the author.

b.
Vacation isn't a blackout.

A teammate logs off for a week and their last wrap keeps answering. No one checks Slack from the beach.

c.
Handoffs compound.

Every published wrap joins a longer record. New hires read what senior engineers read. Context ages well when it's written down.

Built for teams that ship.