Shadow a live CRM with typed decisions

01 · Role in the system

What Jev does here

The inbox receives WhatsApp, Telegram and email messages that a chat model currently answers in prose software cannot act on. The shadow layer replaces generation with one typed request: intencion and division as choices, quiere_humano and lead_escolar as probabilities, urgencia as a 0-3 score. Every answer returns per-option probabilities and a confidence so the caller can set thresholds - act on its own above, escalate below. The layer only decides and logs; it never sends anything into production. A second backend forces any chat model into the same typed JSON shape, so the layer runs before Jev access exists and the difference is measurable on real traffic with one flag. Example runs decide eight messages in about 670 to 700 milliseconds each with confidences from 0.99 to 1.0 on clear intents and a deliberate 0.80 human-handover probability on an ambiguous support message.

02 · Control boundary

Where Jev sits

One batched typed request per message beside the live pipeline, logged for comparison; thresholds in code decide act versus escalate, and a forced-JSON chat backend enables A/B measurement.

Code owns the loop, permissions, thresholds, validation, and side effects. Jev owns only the bounded judgments described above.

03 · Known limits

What this evidence does not prove

• A shadow layer for one small agency CRM, run on example messages in the repository, not a fleet measurement.
• Deciding on inbox text means message content reaches the API when the layer is enabled; the design keeps production sends untouched but does not change that data flow.
• Latency and confidence figures are from the author's example runs.

04 · Attribution

Public sources

This is a Community record: the project was published by a third-party community author.

• alonsuare ↗Community · github · public · checked 2026-09-23