Engineering Constraints

Constraint overview

Area	Constraint	Design response
Files	Five files, 10 MB each	Validate before processing
Long documents	No page-count limit	Process every page in bounded work
AI cost	Free-only	Pause on quota and avoid duplicates
Dates	Evidence required	Deterministic calculation and provenance
Privacy	Private documents	Scoped server access
Email	Multi-day and duplicate events	Durable state and idempotency
Calendar	Portable file output	Deterministic ICS
Retention	15 days from creation	Scheduled cleanup

These constraints keep the prototype affordable and protect user data. They also create visible tradeoffs in processing time, recovery, and long-term access.

File, model, budget, and runtime limits

Accept up to five English PDF or supported image files per submission, each no larger than 10,000,000 bytes. There is no page-count limit, so long documents need bounded page processing with explicit coverage.

Gemini is accessed through server-side credentials and a configured model choice. Structured responses require application validation. Native PDF processing is the preferred MVP approach.

Provider use remains free-only, with no automatic paid fallback. Quota exhaustion pauses work and preserves state. The web host must support server routes and bounded worker execution, and scheduled background calls are needed for email work and retention cleanup.

Date and calendar rules

Installation date is optional. A month and year uses the 15th and remains estimated. Explicit manufacturer triggers take priority. Warranty periods, condition-only actions, ongoing tasks, and short operating durations do not create maintenance dates.

Calendar output is deterministic. Events are all-day, use an exclusive next-day end, and request a 14-day reminder. Only valid eligible dates enter the artifact.

Privacy, storage, and security

Use private storage, server-side workspace authorization, hashed access capabilities, and server-only provider credentials. Treat manuals and email content as untrusted input.

Durable data changes use migrations and preserve ownership, evidence provenance, versioning, and deletion relationships. Privileged operations stay on the server.

Never place backend secrets in browser-visible configuration.
Never log document contents, credentials, or private links.
Never let model text select a workspace, recipient, or authorization scope.
Verify provider signatures and signed reply context.
Keep deletion and retention work narrowly scoped.

Reliability and operations

Use durable jobs, leases, checkpoints, idempotency, and bounded retries. Expensive AI work should run conservatively while deterministic and delivery work can continue.

Free-tier limits may increase latency and temporary failures. The interface needs honest waiting and retry states. Partial page success may remain useful when the missing coverage is clear.

The release checklist must verify provider authorization, background scheduling, retention cleanup, domain configuration, and private storage in the target environment.

Unsupported behavior and future review

The MVP excludes live device monitoring, direct calendar synchronization, public document URLs, password accounts, paid AI fallback, and automatic web result editing.

Before changing a limit or provider, review privacy, cost, migration needs, user messaging, failure recovery, test coverage, and documentation.