Skip to content
Analytics model to flag historical PIDs related to the Suraksha list updateQC optimization to free up agents for old inflow cleanupRefining inflow for manual QCAdding a confidence score for brand flaggings to be used as a lever to QC / skipCurrently the DS models and manual QC is indexed towards new inflow leading to ~25% of escalations which come from existing inflowWe have added 21 agents to run the current ops tasks and the focus of QC is mostly on the new inflow. We are trying to optimize the inflow so that agents freed up can be utilized for old inflow clean up. ()Refining inflow for QC - eliminating SSCATs and alternate keywords leading to high FPsAnalytics model to do OCR, Keyword matches, image matches on all live PIDs which get mapped to a new Suraksha additionConfidence score will be provided by DS for all the image based models. For the text based models, source based precisions will be used as confidence scoreBrand & illegal repeat identification and catalog upload blockUndertaking process to be set up and aligned category KAMsSelective price edit enablement (V2)55-60% of the leakage is coming from the repeated sellersFilter out for P0 suppliers who are 100% fake in nature with thresholds on frequency, # & % of deactivations on pareto basis covering atleast 90% of brand infringements. (~6K P0 suppliers, ~51 lakh live listings )Check for the false positives in P0 identification and minimize the sameExplore manual QC of P0 suppliers overall listings (~51 lakh live listings) to avoid any leakagesTBD - Optimization of the list is yet to be pickedChecking for view share of non compliant listings across major real estates on platformIdentifying the reasons for leakage and sending providing feedback to DS team, manual QC teamView dipstick and instrumentation on adminCalculating the # of DoD non compliant listings across search, PDP, FY feed (REs contribute to 90+% of views)Checking the view share these listings gathered on platform before deactivationIdentifying the reason for leakage and providing feedback for DS retrainingCurrent inflow : 1.7L PIDs, agents: 72 (required : 85)Minimize agent errors by audit and possibility of revising agents targetsAlign on targets for the error rates for agentsProviding continuous feedback and training for low performersCurrently 24% of escalations are coming from agent errors over the last 3 months so we are aiming to make the QC process more efficient to avoid any further leakages ()
Monthly QC of all listings on platform (comes at a cost of ~8000 dollars/run)Optimizing for manual QC bandwidthFrom the last run the inflow count is ~30 lakh PIDs which needed ~52 agents for 15 days to QC which is not feasible so inflow optimization to be figuredCompliance gating v1 - Keyword basedSuraksha list cleanup - Business + Product KRCompliance gating v2 - Image basedDirect deactivations from DS output via tech eventsStrengthening DS feedback loop from view dipstick process~20% of view share leakage is from DS identified cases not getting deactivatedTo accurately track the NS and move some manual dependency to techAlso setup instrumentation for DS feedback loopExtending compliance gating v1 to all edit flows (Seller self serve + Admin)
Alternative keyword coverage is not comprehensiveSellers find out new means using special charactersFixing the precision and recall metrics for all DS modelsIntegrating manual QC / reactivation feedback loops to improve flagging accuracyImproving model recall to 90% and precision to 75%Adding a confidence score for brand flaggings to be used as a lever to QC / skipSetting up admin panel RBAC for core cataloging flows - Deactivations / Reactivations and hidden collection tagging - DoneFill rate exercise done for MRP / COO / Manufacturer detailsTech and product support in setting up these attributes and driving on panel nudges for seller’s action - DoneMRP - To be picked for solutioningAny other attribute - To be mapped from legalUse seller signature to save Seller - Meesho agreement
Core Metric: Non-Compliant Branded Listings Weekly View ShareThe view-based dipstick process has been scaled up from mall brands to the complete Suraksha list Apr 15 onwards. Spike in the view-share trend (NorthStar) and number of deactivated products can be attributed to the increased coverage.Calculation methodology:View share overall % = # total views of the leaked PIDS/#total platform views
Core Metric: Non-Compliant Illegal Listings Weekly View Share
There was a spike in May Week 1 due to addition of ~700 new drugs/illegal classes/keywords. Illegal view share is not stabilized similar to brand due to frequent bulk additions and the nature of keywords. (Eg: money, sex, ling, knife, ipl etc)Calculation methodology:View share overall % = # total views of the leaked PIDS/#total platform views
6M view and KRs
Process Flow
image.png
Compliance KRs 4
Problem arm
Sub KRs
Status
Priority
Details
Tentative timeline - WIP
Required manpower
Primary owner
Old inflow cleanup: New brand/illegal class additions , brand state change cleanup
To be picked
P0
June Week 3
Business
DS
Repeat offence treatment
In progress
P0
June Week 1
Business
Product
View based dipstick
In progress
BAU
W4 June
13 agents
Business
Ops agent accuracy improvements
In progress
BAU
Business
Old flow: Monthly live listings cleanup
To be picked
P2
June Week 2/3
TBD (Check for overlap and keyword stuffing impact)
Business
Inflow guardrail
Tech execution
P0
~5% view dipstick leakage from new inflow
23rd Jun
Product
Inflow guardrail
To be picked
P1
Same as above
Product
Process leakage reduction - Automation
Tech execution
P0
8th Jun
Product
Edit guardrail
To be picked
P1
Cases where sellers edit and add fake images for old listings, E.g - Driddle escalation for Mamaearth
Product
Inflow guardrail
Hygiene - Not allowing special characters in upload and edit flows
Tech execution
P2
Esclataions like Ray+Bin / M@M@ E@rth are currently not detected since -
~21st May
Validation failure lift? and listings drop??
LBYS DROP - Check metric
Product
Better coverage / False positive reduction
Done
P0
Recall stands at ~90% while precision stands at ~75% as per DS, Ops + BAs needs to update the computation for revising the reported numbers
Currently the feedback from manual qc comes with a lag of 12 days. Need the feedback to be as near real time as possible
TBD
DS
Selective QC support (to reduce manual OC effort)
To be picked
P0
Confidence score will be provided for all the image based models. For the text based models, source based precisions will be used as confidence score
TBD
DS
Admin fraud aversion
Partially done
P0
Legal
Ensuring completeness of critical compliance attributes
Partially done
P0
June Week 1
Legal
Ensuring correctness of critical compliance details
To be picked
P1
TBA
Legal
Seller agreement instrumentation
To be picked
P1
Legal
There are no rows in this table
Kaizen metrics:
image.png
image.png
Want to print your doc?
This is not the way.
This is not the way.
Try clicking the ··· in the right corner or using a keyboard shortcut (
CtrlP
) instead.