Zach Zayac

Projects / WITAN

Agentic Calendar & Scheduling

Live · in daily use

Plain-language calendar and to-do control with guardrails that make wrong moves hard.

Updated

What problem it solves

Calendar apps are excellent filing cabinets and terrible listeners. “Move the meeting with R. from 4–5 today to Saturday at 9,” “the Power BI ones should be filed under work,” “delete those and finish the reading one” — real requests are compound, referential, and messy. Most assistants either refuse them or, worse, guess — and a scheduling system that guesses wrong even occasionally loses your trust permanently. This segment exists to take natural, compound, badly-worded asks and execute exactly the right structured operations — or ask exactly the right question before doing anything.

The thesis

Robustness doesn’t come from a bigger model; it comes from decomposition plus grounding. Understand the ask by meaning (not keywords), split it into typed operations, ground every operation in the real assets it refers to, act on what’s certain, confirm what isn’t — and make everything undoable. Each piece of that sentence is its own small, testable component.

The ask pipeline A plain-language ask is classified by meaning, parsed into structured operations, grounded against real assets, then acted on with undo, confirmed, or refused. ONE PIPELINE · EVERY STEP TRACED · EVERYTHING UNDOABLE The ask "move the 4 o'clock to Saturday, and file the BI ones under work" Switchboard classifies by meaning, not leading verbs — then splits compound asks Event parser To-do / modify parser Move parser narrow on purpose — each one testable alone Grounding gate every operation resolved against the real calendar and to-do assets certain → act, with undo uncertain → ask first known-wrong → refuse
The ask pipeline — classify by meaning, parse into typed operations, ground against reality, and only then act. Wrong moves die at the gate.

Key features

  • A meaning-based switchboard. The orchestrator classifies every ask by what it means, not its leading verb — “those can be removed” routes as a modify, any-word-order reschedules route as moves — then dispatches to the right specialized parser.
  • Specialized parsers, kept narrow on purpose. One parser turns compound to-do asks into structured operation lists; a sibling handles event creation (ranges like “4 to 5”, meeting defaults); another handles moves. Keeping them narrow keeps each one testable; the switchboard above them does the routing.
  • The grouped-ask spine. Bulk requests flow through one reusable pipeline: split the ask → resolve the referenced set against reality (confident / excluded / ambiguous) → a relatedness agent judges which unlabeled items actually belong to the group’s meaning → execute with per-item confirms and undo. New bulk operations reuse the spine instead of reinventing it — and every hop in it is a Cortex-traced event.
  • A grounding gate that refuses. Before anything acts, a catalog-grounded check refuses known-wrong operations outright — “complete the event” when the target is a to-do gets asked about, never guessed. Adversarial tests pin this boundary.
  • Time-awareness features on top: a month view whose multi-day bands are live condition objects (that machinery has its own page), an adaptive gap-filler that proposes uses for the week’s empty stretches (one batched card, propose-only, a dismissed week stays dismissed), a deadline negotiator that ranks which soft deadlines to move when capacity overflows, and completion-relatedness — finishing a to-do surfaces the open items that obviously pair with it.
  • Write-layer invariants: duration floors, title normalization, timezone correctness — the boring guarantees that make everything above them trustworthy.

The week, auto-filled

The gap-filler is the segment’s most agentic behavior, so its restraint matters most. It scans the week ahead for large empty stretches, ranks candidate uses from overdue work and top priorities across every life domain, and delivers one batched proposal card — not a drip of nags. Approving places the blocks; each placed block carries a plain-language receipt of why it exists and a release condition (“if the driving task finishes early, this time goes back”). A dismissed week stays dismissed.

Many of those empty stretches don’t appear by accident — they’re freed upstream, when a trip ends early or a condition releases its windows. That handoff comes from the Predictive Conflict Detection & Cascade engine, which is the other half of this story.

The week, auto-filled — illustrative render A five-day week where solid blocks are committed time and dashed blocks are the gap-filler's proposals, each carrying a plain-language reason. This week adaptive gap-filler · propose-only · placeholder data MON TUE WED THU FRI Class Training Standup Deep work — proposed — Call Field day Study block — proposed — Coaching 1:1 PT test Prep — proposed "Deep work, Tue 10–12 — why: thesis review overdue · releases the time back if the driving task finishes early"
Illustrative render of the week view (placeholder data): solid blocks are committed, dashed blocks are the gap-filler's ranked proposals. One batched card per week; every placed block carries its "why" and a release condition.

Highlights

  • “Move the meeting from 4–5 today to Saturday 9am” moves the existing event — the naive failure (create a duplicate, leave the original) is structurally prevented, not patched.
  • The gap-filler closed its full loop — noticed empty time, proposed ranked fills with plain-language receipts, placed approved blocks — on its first live day.
  • The wrong-asset failure class (acting on a to-do when you meant an event) was eliminated as a class, and a test battery keeps it dead.

Development log

July 8, 2026

Page created. State as written above: switchboard + parsers + grouped-ask spine + grounding gate all live and in daily use; gap-filler and deadline negotiator shipped this past week; month view with date-first layout and awareness bands landed alongside.