PAPERFLITE

Start with the document someone already has

A bank can have 90,000 documents and still leave someone stuck if they don’t know what to search for. At Paperflite I started with the document in front of the reader. I built a recommender that extracted keywords and used cosine similarity to find related material in the collection of a public-sector bank client in Africa. That gave readers a route into the library without needing to know the exact title of the next document.

Let usage inform the ranking

I also built a separate stylometric engine that grouped documents by writing style using K-means. After the recommender launched, I used readers’ usage history to re-rank its results. The initial ranking came from the documents; the next iteration could also draw on how people used them.

I was the Data Scientist and ML Engineer on this work. Alongside the recommendation engines, I produced weekly Tableau reports on feature usage, regional traffic and time spent by feature and client. Product, engineering, marketing and sales could look at how the platform was used alongside their own questions about it.

The work covered the first recommender and its post-launch iteration. The 90,000+ figure is the size of the document collection.