What Is Sivella?

Clara began as an attempt to reduce the tiredness of carrying a household in one person's head. It became Sivella as tasks, calendars, money, meals, decisions, and memory grew into one connected system.

Sivella is the household operating system I am building to share that load. It gives couples and families one place for tasks, decisions, money, calendar availability, kitchen planning, reminders, energy, and household memory. A conversational assistant can read the relevant context and help the household act on it.

The product should let someone record a fact or get an answer in a few seconds, then leave. A home does not need another dashboard to maintain.

Today shows what needs attention

The Today view reduces the household’s open work into a smaller daily picture. It can show a primary next action, current priorities, budget pressure, energy state, reminders, and unresolved decisions.

A quick brain dump captures a thought before it disappears. Sivella can turn that thought into a task, park it for later, or connect it to an existing thread. Simplify mode reduces the amount shown when the household has less capacity.

The purpose of the screen is relief. It should answer “what needs us now?” without presenting every unfinished item at once.

Tasks carry more than a checkbox

Household tasks can have an owner, due date, energy requirement, estimated time, cost, recurrence, subtasks, and shortlisted options. Errands and repairs remain distinguishable from ordinary tasks because they behave differently in real life.

Sivella can suggest a time by checking calendar availability. It can suggest an owner by combining workload with free time. Reassigning a task clears an old schedule slot so the new owner does not inherit a time chosen for someone else.

The household stays in control. Suggestions explain their reasoning and can be changed.

The assignment model combines availability with workload instead of treating task counts as capacity.

Calendar access protects work details

Sivella can connect Google, Outlook, Apple, and CalDAV calendars through read-only access. It stores free and busy blocks, not meeting titles, attendee lists, or work notes. That is enough to avoid suggesting a 3 PM errand to someone already occupied from 3 to 5. Access is logged, and the connection can be removed.

Money gives decisions context

The Money area supports household and personal ledgers, transactions, expense accounts, category budgets, monthly limits, CSV import, and exports.

Budget goals and upcoming expenses make the view forward-looking. Sivella can compare a possible purchase with current spending, scheduled expenses, and the household’s goals. It returns an affordability assessment as advice. The household still decides.

Personal ledgers stay separate from household context. The assistant uses shared financial data when helping with household questions.

Kitchen connects groceries to meals

The Kitchen area combines shopping lists, pantry items, meal preferences, and meal suggestions. A household can record what it has, add what it needs, group items into categories, and use rough ingredient lists to find meals that fit.

Onboarding learns dietary preferences and shared meal choices. Household members can have overrides where their needs differ. Custom meals can be added from pasted recipes, with ingredients normalized into pantry items.

The longer-term value comes from connecting restock patterns, selected meals, and the household budget. The current product builds the pantry and meal foundation for that work.

Memory keeps the threads connected

Sivella stores household context with visibility rules. Memories can be private, shared with the household, or limited to the right people. Members can inspect and edit stored memories instead of treating the assistant’s memory as a hidden system.

The context engine assembles only what a question needs: tasks for planning questions, budget data for affordability, availability for scheduling, and relevant past decisions when the household revisits a topic.

This keeps the assistant focused and limits how much personal information enters each model call.

Sivella has a personality system

Households do not all want the same tone or amount of detail. Sivella’s personality modes change the interface, density, and voice around the same underlying household data. Serenity and Focus are currently live. Nurture, Steward, and Mirror are designed but remain disabled until their full behavior and surfaces are ready.

The personalities should change how help feels without changing budget math, permissions, or task rules.

Where Sivella is still going

Prediction, fairness, and proactive help are still active work. Pantry history should eventually improve restock timing. Workload suggestions need to learn without repeating an unfair division of labor. Reminders need enough context to help without turning the home into a queue of alerts.

The product is further along than the Clara mockups in May, but the test has stayed the same: after Sivella helps, the household should have less to remember and less software to maintain.

Building with AI: I design the household boundaries and review the complete state transitions; AI writes much of each Django app and frontend surface. That makes separate modules appear connected before they truly are. The kitchen loop and task-owner invariant now prove the connection by checking that one action changes the next part of the household system, not merely the card on the current screen.

How the surfaces stay connected

These surfaces are not one large assistant prompt. They are separate Django apps and services with explicit contracts. Tasks use scheduled_start and scheduled_end; availability exposes free/busy windows; budget context reads ledger and upcoming-expense data; memory applies visibility rules before building model context.

I walked the Kitchen loop through the running Django stack rather than a mocked screen:

meal tiles → diet profile → preference pick → seed pantry
→ mark Loose Tea out → auto-add to shopping list
→ mark bought → pantry returns to in_stock

That run created 54 pantry items, produced the expected source=auto_out shopping-list entry, and completed every HTTP step successfully. It showed the value of the product's connected model: “bought” was not a decorative checkbox; it changed pantry state.

I also followed task placement across availability, assignment, and the assistant. Reassigning a task cleared its old schedule, free-time suggestions respected privacy-scoped availability, and adjacent budget work did not change those rules.

Today, Tasks, Calendar, Money, Kitchen, and Memory now share state transitions instead of merely linking to one another. Ownership and privacy invariants sit in the services, while live HTTP walkthroughs cover loops where one action must update another part of the home.