Command
Growth AI
The model layer behind Aiera. Every enquiry is scored the moment it lands, routed to the right channel, and re-scored after each conversation.
Enquiries scored
15,200
+18% QoQRolling 90 days
Score → visit lift
2.7×
vs unscoredTop decile against baseline
Model precision
0.81
Hot band, measured on booked outcomes
Median score latency
4.2s
−1.8sLead created to score written
Weekly precision of the hot band
Share of hot-scored enquiries that reached a site visit within fourteen days
The step at week nine is the Hinglish transcript feature going live — intent read from what the buyer actually says, not only what they clicked.
Enquiries by score band
Rolling 90 days
Junk is suppressed before it reaches an advisor. That alone returns roughly 90 advisor hours a month.
Feature weight in the current model
Relative contribution to the booking probability
Routing outcomes
What the model decided to do with each enquiry
Hot and warm bands, called inside nine minutes
Cold band with a real project interest
Ticket above ₹3 Cr, always human first
Duplicate, test or invalid number
Model versions in production
| Version | Live since | Scope | Precision | Recall | Note |
|---|---|---|---|---|---|
| ir-score-3.2 | 12 Feb | All enquiries | 0.81 | 0.74 | Adds Hinglish transcript intent |
| ir-score-3.1 | 04 Jan | All enquiries | 0.76 | 0.71 | Held as fallback |
| ir-nri-1.4 | 22 Jan | NRI enquiries | 0.79 | 0.66 | Timezone-aware call windows |
| ir-resale-0.9 | 01 Mar | Resale intent | 0.68 | 0.58 | Shadow mode, not yet routing |
What the model is asking for
Precision in Mumbai trails the book by six points because the outcome field is left blank after site visits.
Sales ops · this week
Shadow accuracy has held above 0.66 for three weeks. Routing resale intent to the dedicated desk should lift the band.
Growth · next sprint
No traffic in twenty-one days. Keeping it warm costs nothing but muddies the audit trail.
Platform · anytime
Honest limits
The model reads intent, not affordability
A buyer can sound ready and still fail a loan check. Pre-approval is a separate signal and it is deliberately weighted low until the bank responds.
Precision is measured on booked outcomes only
Enquiries that are still open are excluded, which flatters recent weeks slightly. Read the twelve-week trend, not the last point.
How to read this page
Command · Growth AI
Window
Rolling 90 days of the India Realty pilot, refreshed nightly at 02:00 IST.
Sources
Aiera call logs, WhatsApp Business API, CRM stages, Meta and portal spend exports.
Joins
Every number is joined to a booking where one exists, not to leads, so cost lines stay comparable.
Limits
Ninety days is short for a ₹1.5 Cr purchase. Late-window cohorts are still open, so treat conversion as a floor.