01 — The ProblemOne report, five backend systems

When a fleet partner reports a vehicle isn't reporting telemetry, the honest answer could live in any of five or six different places — consent status, enrollment state, the telemetry push history, the odometer feed, network reachability, SIM and carrier status, or a handful of task and event tables. Manually, that's a slow crawl through the CRM and several backend systems before an analyst can even start writing the case note.

The notebook exists to make that crawl unnecessary for the majority of cases, and fast for the rest.

02 — Tier 1A deterministic gate function, not a checklist

A single input cell sets shared context — one or more VINs, the case number, the analyst's name — and auto-fetches case history from the CRM. Every downstream cell reads from these same variables, so the notebook behaves statefully across cells instead of each cell re-deriving context on its own.

Tier 1 runs five independent queries — consent, enrollment, telemetry push history, odometer, and network reachability — merges them into a single per-VIN DataFrame, and feeds that into a gate-evaluation function. The function walks a fixed decision tree: consent, then account and vehicle gates, then unit state, then connectivity agreement, then FAN mapping, then push blockers, then push history. It emits a verdict, an exact issue code, and a recommended next step. Most cases resolve here, without touching anything downstream.

A gate function that always walks the same tree in the same order is what makes the tool trustworthy. An analyst can predict what it'll say before it says it — and check its work.

03 — Tier 2Nine cells, run selectively

When Tier 1's verdict doesn't fully resolve the case, nine independent cells cover the specific failure domains that Tier 1 can flag but not fully diagnose: enrollment task state, SIM and carrier status, transaction and command history, batch remediation history, full telemetry task status, GPS issue flags, enrollment event trace, connection event history, and plan ID cross-reference.

Each cell is scoped to one failure domain and is only run when Tier 1's verdict points there — the notebook isn't meant to be run top to bottom every time. That selectivity is what keeps a nine-cell deep dive from turning into a nine-cell ritual.

04 — Tier 3Turning diagnostics into a case note

Once a verdict exists — whether from Tier 1 alone or after a Tier 2 deep dive — an LLM call takes every diagnostic variable collected so far, plus the Siebel case context, and produces a ready-to-send partner-facing case note. The prompt enforces a strict output shape: verdict, findings, recommended next steps, and — for any opt-in or opt-out action lines — the correct partner identity and app ID, so the generated note can't drift on those specifics even if the surrounding language varies.

05 — Driven by GenieFrom a pasted case number to a finished note

The whole flow is designed to be fully driven by Genie: an analyst pastes a case number, VIN, and customer email into chat, and the assistant sets the notebook's widgets, runs the appropriate cells for that case, and generates the case note automatically. The analyst's job becomes reviewing the output rather than assembling the diagnostic path by hand.

→ Compute

Everything runs on Serverless CPU using Python and Spark SQL only — no Scala, no R — which keeps the notebook portable and avoids any cluster-specific dependencies for what's meant to be a frontline, high-frequency tool.

06 — TakeawaysTiering is a UX decision as much as an architecture one

  • Put the deterministic logic first. A fixed decision tree that resolves most cases builds trust before any AI-generated content ever appears.
  • Scope deep-dive cells narrowly. Nine cells that each answer one question beat one cell that tries to answer nine.
  • Let the AI layer generate the note, not the diagnosis. The verdict comes from deterministic logic; the LLM's job is turning that verdict into readable, correctly-formatted partner communication.
  • Chat-driven execution is a real UX win when the underlying tool is well-tiered — Genie can only "just run the right cells" because the notebook already knows which cells are right.