Build Reliable Employee-Operations and Expense Intelligence Systems

# Build Reliable Employee-Operations and Expense Intelligence Systems Design four connected capabilities: a tagged event counter for dashboards and billing, a driver-pay ledger, an employee-termination orchestrator across external systems, and an LLM-backed expense assistant over employee and reimbursement records. ### Constraints & Assumptions - Mobile clients can upload a large offline backlog; events may be late or duplicated and cannot be silently lost. - Dashboards tolerate tens of minutes of freshness delay, but monthly billing must reconcile exactly. - Some termination actions are parallelizable, some have prerequisites, and some cannot be safely repeated. - The assistant must enforce company and employee data boundaries before model access. ### Clarifying Questions to Ask - Which event fields form the billing identity and which are mutable tags? - Does a pay-rate change affect only future work or also unsettled historical work? - Which termination systems support idempotency, webhooks, or compensating actions? ### Part 1 — Event counter Design ingestion, deduplication, tagged aggregation, dashboard serving, and exact monthly billing for mobile and web SDK events. #### What This Part Should Cover - Durable acceptance and idempotency - Event-time windows and late-data policy - A reconciled billing ledger separate from approximate dashboards ### Part 2 — Driver pay Model drivers, rate history, completed jobs, overlapping time spans, total accrued cost, and paid versus unpaid amounts without rewriting history. #### What This Part Should Cover - Immutable work records and effective-dated rates - Overlap and duplicate semantics - Accrual, payment, and reconciliation APIs ### Part 3 — Termination workflow Orchestrate seven to ten dependent revocation and payout actions, minimize completion time, and resume automatically after interruption. #### What This Part Should Cover - A persisted DAG or saga state machine - Idempotency and compensation per step - Priority, retry, webhook, and operator escalation behavior ### Part 4 — Expense assistant Turn a natural-language expense question into safe structured operations over two large JSON-backed datasets, support aggregates and streaming, and explain when to use deterministic tools versus an agent. #### What This Part Should Cover - Tenant-scoped parsing and authorization - Typed query plans and aggregate correctness - Streaming UX without leaking unvalidated partial results ### What a Strong Answer Covers - Clear system boundaries - Recovery paths for every partial failure - Privacy and correctness checks before optimization ```hint Separate approximate serving from exact ledgers Dashboards can use compacted aggregates, but payroll, billing, and access revocation need durable state, explicit identities, and replayable evidence. ``` ### Follow-up Questions - How would a 100-fold event increase change the architecture? - How would you prove that a termination workflow is complete?