Gurpreet Singh Padam

Relationship Managers (RMs) were uploading incorrect or random photos during loan applications instead of the actual customer's picture.

• CPU team received mismatched photo submissions
• Had to raise queries back to field RMs to get correct photos
• Each query loop added 1–1.5 days to TAT
• CPU bandwidth wasted on back-and-forth communication
• Data quality deteriorated across the loan file

Integrated a Face Match API directly into the LOS at the point of application submission.

• RM captures a live photo of the customer at the time of visit
• RM uploads the customer's Aadhaar card (photo ID)
• API compares live photo vs Aadhaar photo — returns a match % score
• If score > threshold (60%) → case can be submitted
• If score ≤ threshold → submission is blocked, RM must retake photo
• Zero CPU involvement required for photo validation

Automated gate eliminated the problem at source — before it ever reached the CPU team.

• TAT reduced by 1–1.5 days per case
• CPU query volume for photo issues dropped to near zero
• RM accountability improved — they must get the right photo
• Audit trail created — match % logged per case
• Data accuracy improved across the loan file

CAC was high because campaigns were targeting new cold users indiscriminately, ignoring an existing warm user base that hadn't transacted yet.

• No differentiation between previous users and new leads
• Same channel and frequency used for all segments
• Budget wasted on users unlikely to convert
• No model to predict who would sell their car next

Built a two-layer data model using IV (Information Value) for feature selection and XGBoost for propensity scoring.

• Identified previous users who hadn't yet sold their car to Cars24
• Used IV to select high-signal features (car age, past interaction, location, etc.)
• XGBoost assigned a propensity score to each user
• Users bucketed into cohorts by score band
• For each cohort: determined optimal channel (WhatsApp / RCS / SMS) and frequency
• High-score cohorts → WhatsApp (rich, immediate)
• Mid-score → RCS, lower score → SMS with lower frequency

Precision targeting dramatically improved campaign efficiency vs default outreach.

• CAC reduced by 15–20% vs new customer or unmodeled outreach
• Targeting accuracy improved by 20–30%
• Communication frequency optimised — reduced noise for low-intent users
• Model became a repeatable engine, refreshed monthly
• WhatsApp open rate significantly higher for high-intent cohort

All leads were routed to a First Insights team for initial qualification before handing off to the ground team at properties.

• Agents spent majority of time on basic questions: stay duration, budget, preferred location
• High volume → agents couldn't focus on complex or high-intent leads
• Delay between lead creation and first contact increased drop-off
• Ground team at properties received unqualified, incomplete handoffs
• No standardised data structure across lead types

Implemented a WhatsApp chatbot flow that replaced the entire initial qualification stage.

• Lead created → automated WhatsApp message sent instantly
• User completes qualification journey via chatbot: budget, duration, city, property type
• Chatbot shortlists and presents matching properties
• User selects preferred property from options
• User picks a date and time for their property visit
• Only after visit confirmation — notification sent directly to property staff who will assist them
• First Insights team freed up for complex / high-value cases only

Automation replaced repetitive human work, improving both speed and quality of the lead journey.

• Sales team dependency reduced by 60%
• Lead-to-booking conversion improved by 20%
• Time-to-first-contact dropped from hours → seconds
• Property staff received fully qualified, scheduled visits — no surprises
• Standardised data captured for every lead through the chatbot
• First Insights team refocused on high-value edge cases