Jan 13, 2021
The Pivot
The majority of the 12 hours spent at the marketplace went to talking to Kirana store owners. Out of three Kirana stores, two were unorganized & one was organized. The organized one uses Marg & Tally software for Inventory management & Taxation records maintenance respectively. The billing part is done by a barcode reader. The setup as a whole seems very efficient at the first look. But there is a missing automated piece that reduces the effectiveness of this setup, Manual Data Input of Purchase Invoices.
Whenever the delivery from the wholesaler/distributor comes to the store, a printed purchase invoice/receipt is handed to the store along with the goods. Unorganized stores only use them for tax filing purposes. But the organized stores input the receipt data into their inventory management software manually. This process is done stepwise:
1. The item to be entered is scanned using a barcode reader. If item mapping fails, then the item is manually selected/entered.
2. Item details like Quantity, HSN Code, Discount, Rate, Tax, etc. are entered manually against the scanned item.
On average, an organized Kirana store receives about 90 invoices with 20 item lines per invoice weekly. The time to enter 1 item line ranges from 30 seconds to 1 minute. The average time spent on entering invoice data takes about 15–20 hours weekly. The bigger the store operation, the more time to input invoice data. Big stores usually hire people to carry out this process incurring Rs. 10k/ monthly on average. These numbers are not just picked from an article/study from the internet but the result of the customer development process I ran which constituted talking to 60+ Kirana Stores & spending over 100 hours.
The challenge now was to bring down the time & effort spent on this redundant process affordably. The idea we pivoted to was to make an integration between invoice images & inventory management software. We decided on building an AI-powered OCR-based software that takes in invoice images & returns a customized template-based excel file that could be imported to the inventory management software. This would reduce the time to input data from hours to seconds & if we could hit our theoretical cost projections, we could end up providing a value-for-money way to our users.
I decided to validate if this problem was as big as we thought & designed a user interview process. The process included a questionnaire that had indirect questions that would be easy to understand from the store’s perspective followed by showing the prototype of our product to give them a complete picture. I created a prototype using Balsamiq to sketch low-fidelity wireframes & Marvel to wire them together. I set a target of interviewing at least 50 organized stores before concluding anything.
Prototype Video: https://youtu.be/29QGjdcvmV0
In the meantime, we got selected in a top VCs-led initiative PayItForward & a Lightspeed Ventures India partner took us under their mentorship. We expanded to 7 members adding people proficient in AI/ML. The tech members started with the initial development of the prototype & I continued my user interviews.
Fast forward three weeks, I completed my user interviews & the tech team was ready with the first draft of the algorithm to convert images to customized excel files. The result of the user interviews was pretty overwhelming. The problem we were trying to solve was absolutely present. About 83% of the stores were interested in trying our product asap & would pay anything less than Rs. 1k/month if the results were as conveyed. Three stores were so convinced, they told me to make them our genie pick i.e they would help us solve this problem in every possible way.
The next step was testing our algorithm’s first draft. We took 600+ images of the purchase invoices to train the model & get the initial results of the algorithm. We gave early access to these three customers and served them for more than two weeks by passing their invoices through the model and using human input to rectify mistakes; returning them the converts on daily basis.
The results of the first draft testing were what changed everything. We expected to get a 95%+ accuracy but even after using different ML models, image processing techniques we got only 73% accuracy. The worse thing was that the algorithm missed the item lines randomly. The accuracy for a full item line was even lower at 43% on average. The team felt the low accuracy was partly because of the low sample size, so we added another 600 invoice images to the sample. The results were contrasting & the overall accuracy decreased to 68% with the item line accuracy at 27%. We had failed in developing an image to excel algorithm. We had anticipated that the challenge would be difficult as Microsoft Excel with its image to excel feature cannot give anything close to 90% accuracy.
We then moved to see if there is an outsourcing solution to this technical problem. We came across one website that offered API support & whose results were good & with a final human touch, we could get the image to excel process completed in a short time. But this option turned out to be extremely expensive. The cost projections would shoot up by 200% following this path making it financially unviable.
The underlying value behind doing this process was getting the purchase data & its frequency that would enable us to get the inventory of the Kirana stores online. By running analytics on the purchasing pattern of a Kirana store, we can predict the availability of an item in the store near accurately, getting near-real-time inventory data thereby, taking the store online. This data could very well be the missing piece in the marketplace-based hyperlocal delivery puzzle faced by Grofers, Swiggy, Flipkart, JioMart, etc.
Our mentor is a veteran in the Consumer & SME sector & they helped us verify the value of the data by not only connecting us with upper management of the above firms but also brainstorming throughout the process with us. Using all the learnings, we decided to give another shot to get the purchase data by a different approach. If we could get the digital version of purchase invoices, we would be able to process them in a more efficient & cost-effective way.
The turnaround in a short time had taken a lot of toll on the team. We were all running thin on motivation & time. As it was mid-September, the placement & internship session was at its peak & the clarity on what we would do next was at an all-time low. That is when we collectively decided to take a break from this project & evaluate further options after some time. We thanked our mentor & the store owners for supporting us with their time & energy.
The algorithm we worked on is available here: https://drive.google.com/drive/folders/1LK9Qpdk6gSewGgHHAJFSrM4KMUPma6LK?usp=sharing
Logging in 850+ hours, getting uncomfortable, pitching to people, taking criticism constructively, showing up each day for more than 5 months may not have resulted as expected but it did result in immense learnings that turned out to be far more precious in the hindsight.
