Keep the source intact.
Excel, CSV, manual edits, photos, and PDFs enter a raw layer before the planning model makes them comparable.
The original value remains available for review when a normalized field needs correction.
Selected work / ClassCover
An offline-first planning desk for school substitutions and invigilation. It turns messy timetable inputs and live availability into recommendations a coordinator can inspect, edit, and safely apply.
01 / The problem
School coordinators work from exports, corrections, absence notes, and exam sheets. A useful system has to preserve the raw input, normalize it enough to reason about, and still make the final decision legible to a person.
02 / Product walkthrough
Here is the product behavior worth showing: start with the timetable, mark the absence, inspect candidates, and apply one explainable change. The controls below are a synthetic walkthrough of the ClassCover desk.
Substitution planner / 08:30 start
Excel, CSV, manual edits, photos, and PDFs enter a raw layer before the planning model makes them comparable.
The original value remains available for review when a normalized field needs correction.
Raw rows become teachers, classes, assignments, and relationships without pretending that every source value was clean.
Normalization improves search and planning; it does not erase ambiguity.
Availability, timetable conflicts, release chains, and fairness history shape a shortlist. The coordinator still sees why a candidate appears.
The AI assistant can explain and propose; the product boundary is the review step.
Changes are staged as durable planning selections. Applying a chain makes the operational result explicit and preserves a path back.
A recommendation is useful only when the transition from suggestion to system state is safe.
03 / What I changed
Not an “AI built a timetable” claim. The interesting engineering is the seam between imperfect data, a constrained planner, and a coordinator who needs to stay in control.
04 / The result
The strongest outcome is a calmer operational loop: import what the school already has, understand the affected slots, inspect a recommendation, and apply only what is ready.
Research brief based on the server project, tests, deployment configuration, and the project questionnaire. Diagrams are simplified explanations, not production topology or benchmark telemetry.