Pavel Samsonov

Data Extraction

Designing a human-in-the-loop tagging platform to improve speed, accuracy, and cost of data extraction from financial documents.

    My roles

  • created user research practice by training customer support & sales colleagues in best practices for gathering data on user needs
  • designed and prototyped a new interaction technique that eliminated manual & repetitive steps slowing down the process
  • led a feature team and contributed to front-end development to deliver the new experience within the tight deadline required

    Product impacts

  • 60×
    60× faster extraction through the design of an innovative interaction technique
  • 80%
    80% reduction in operators, reserving human attention for training and edge cases
  • 99%
    99% lower development costs after the need for custom scripting was fully eliminated
  • 0%
    0% file rejection achieved by routing high-ambiguity documents to human operators

At a glance

The scope of work required a deep understanding of two incumbent processes - the bank's manual data extraction work, and Workfusion's existing data extraction UI, developed without user experience design support. I organized a group of developers and customer-facing colleagues to act as a research team for understanding the customer's current needs, and the desired future state.

Modeling the present state
  • Technology leadership
  • Data science team
  • Customer Success team
Defining the desired state
  • Crowdworkers
  • Business stakeholders
  • Data requesters

A deep understanding of the customer journey allowed the team to rapidly test MVP candidates produced by the developers, and iteratively identify gaps and inefficiencies. To enable this process, I partnered with Workfusion's heads of Engineering and Data Science. This allowed the team to understand the underlying platform's full capabilities. I designed an experience that addressed user needs while being feasible to implement within the client's desired time frame. This prototype automated repeated actions, put frequently-accessed information close at hand, and allowed Workfusion's AI to help workers in real time.

Iterative design
  • Sequence diagramming
  • Mockups
  • JS prototypes
  • A/B testing
Go-to-market process
  • Metrics analysis
  • Stakeholder alignment
  • Release and iteration

Even though the proposed scope was beyond the initial plan, the success of the prototype demo convinced the CTO to add more developers to the product team. I worked with them as designer, front-end developer, and product owner to deliver the proposed design on time and to spec. The resultant product would go on to account for over 90% of Workfusion's business at the time.

Tagging UI The UI used to create examples for training the model, and to handle exception cases where the AI expressed low confidence. Human tagging speed improved up to 60x, with no compromise in accuracy.

Get in touch if you want to learn more about this case study.