Weekly Updates

Summary

1. PayList has got launched.
2. All events for payment reconciliation are in place now
3. Various improvements that help in improving login sucess rates have gone live
4. User level plan acess mapping for Select is live

MAU

PL MTD MAU is up slightly compared to previous month: Over long term we have the problem that $MAU = \frac{NU}{1 - r}$ where $NU$ is monthly new users and $r$ is retention. $NU$ is not up to the levels to support existing $MAU$ and $r$ also has deteriorated if anything (recovering post DDD)

Following is the long term graph of monthly PL NU (even first txn in Pi3, user is counted when they do PL txn for the first time)

SELECT
   SUBSTR(first_pl_txn_timestamp, 0, 7) AS mth, COUNT(*)
FROM
   public.user_profile
WHERE
   first_pl_txn_timestamp IS NOT NULL
GROUP BY 1
ORDER BY 1

Separately, Pi3 NU continues to grow -

Delinquency

RU (to date) PL delinq for Dec 2024 to March 2025 have been ranging from 78bps to 81bps. Amazing!

It is trending slightly up as months are passing - early repeat users are becoming slightly more dominant.

NU PL delinq

NU DPD25 PL delinq is stable - increased from several months ago. Will additional delinq be compensated by increased retention? That's a question to which we will seek the answer.

Paylist is launched

Paylist was launched to public with last Wednesday release. Congrats to all for launching it! Here is Narender celebrating the launch.

So far (as on July 27 11 AM), 119 users have set up mandates and 34 users have added bills.

Login improvements

Over past week or so, three changes have gone live that should improve login success rates.

1. In a span of 365 days, no more than 2 phone numbers could login from a given device. Now, we have reduced the period of 365 days to 90 days.
2. In the backup flow, where the user sends us an SMS, in case they enter their phone number as P1, but send SMS from phone number P2, then we will display the last few digits of phone number P2 so that user better understands they need to use the right SIM.
3. Some devices (like xiaomi) don't support login via auto read of OTP. For them we directly skip to send SMS (fallback) mode, than having the user experience failure with the main flow.

At aggregate level, there is no dip in login failure rates - infact it has increased in past few weeks from 10.xx% to 12.xx%. We will see what is going on here.

User level plan access mapping

We built the capability to have variants within select v2. So, now we can have users tagged to select 2 in such a way that they don't get 30 day billing cycle. This way we can expand the select plus plan (you get billbox, priority support, but not 30 day select). We don't want to expand 30 day select till the time we have determined that delinquency for this plan is alright.

Neptune cost reduction

We have reduced neptune cost by around $25 per day. Vinayak had told me what we did, but I am forgetting as of now.

SIKA testing automation

I was quite excited that Moumita is working on SIKA testing automation. Here is a video that she shared - it is all being done code using selenium. This video is good.

Churn Diagnostics

Urmi has finished her churn diagnostic model. We can, with some certainity say that a user is going to turn delinquency, and also why are they going to be delinquent. We will be doing an experiment with remedial actions (remedial actions would be - waiving fines, increasing credit limit a bit, or sending comms

Ensemble in approval model

First ensemble approach is no better than SoTA. Let's see if we can get better results from ensemble.

Rokibul's tech talk

Rokibul gave a tech talk titled Observability Beyond CloudWatch: A Deep dive into Our OTel-Based Stack. Here is video of the same:

Gokwik TPV

Gokwik TPV continues to increase

SELECT 
   bucket_id, 
   ROUND(SUM(IF(product_id = 'a21b31b4-74b8-4a6e-a2a8-28b85d1f8315', txn_amount_in_paise, 0)) / 100) AS pl_tpv,
   ROUND(SUM(IF(product_id = '541eeb1c-8212-40f7-af04-05990ff00961', txn_amount_in_paise, 0)) / 100) AS pi3_tpv,
   ROUND(SUM(txn_amount_in_paise) / 100) AS total_tpv
FROM public.transaction_events
WHERE bucket_id >e; 'date_2025_06_01'
AND gateway_id = 'cf017b96-2adb-4d3f-a115-e3e58867a3ee'
AND event_name = 'TransactionChargedEvent'
GROUP BY 1
ORDER BY 1