https://youtu.be/-_8I7LTJZRE

Context

Industry: E-commerce / distribution with a delivery-driven sales pipeline. Current process: deals moved through packaging, shipping and post-delivery feedback stages, but nobody tracked which ones had stalled. Health checks were manual, inconsistent between managers, and the CEO only saw the pipeline when someone assembled a report by hand.

Problem

• Deals silently stalled between stages — a package sitting at a pickup point for 60 days looked the same in the CRM as one shipped yesterday
• No shared definition of "at risk": every manager judged by feel, so the same deal got different verdicts
• Alert fatigue: any naive alerting repeats the same deals every week until people mute the channel
• The CEO had no aggregate view — only individual deal cards, never the systemic pattern behind them

Solution

• Deal Health Score Analyzer — scores every active deal 1–10 using five weighted criteria (communication 30%, pipeline progress 25%, task execution 20%, interaction quality 15%, deadline health 10%). The score is computed in code, not by the model, so identical data always produces an identical score.
• Activity Sub-workflow — pulls calls, notes and tasks from Zoho in parallel and reports explicitly which entities failed to load, so an API error can never be mistaken for "no activity".
• Alert deduplication — an alert fires only when the category changes or the score drops. Everything else rolls into a single weekly digest with week-over-week deltas.
• Daily Delivery & Feedback Monitor — a separate daily pass over delivery and post-sale feedback stages, with deterministic priority rules (urgent / high / medium / low).
• CEO Weekly Report — aggregates the scored pipeline into a close-rate forecast, three systemic risks with root causes, and a prioritised weekly plan split into one-off fixes and process changes.

Process

• Weekly scoring: Zoho Deals → read previous scores → loop per deal → Activity Sub-workflow → deterministic metrics & score (Code) → AI qualitative analysis → Google Sheets + conditional Telegram alert → weekly digest
• Daily delivery pass: Zoho Deals → split by stage & compute priorities (Code) → AI agent enriches each deal via a Zoho tool → HTML report split across Telegram's message limit
• CEO forecast: Google Sheets → aggregate → AI agent → Telegram

Division of labour between code and model. Everything countable — days since last contact, days in stage, overdue tasks, weighted score, priority, amounts at risk — is computed in JavaScript. The model does only what code cannot: read the tone of notes, spot contradictions between fields, and formulate the next action with evidence quoted from the source record.

Reliability engineering