Viko

Cloud based (PaaS) text mining platform.

Viko

Industry: Technology
Size: SME
Location: London, UK
Website: www.viko.co.uk

Viko uses artificial intelligence to provide direct answers to stakeholder questions e.g. "what is the check-out time?" or "when are bins collected". We're fine-tuning NLP models for a range of sectors and use cases including hospitality, local government and finance (insurance).

Tech used

Python
Tensorflow
JS/Node

Features

Question answering
Semantic search (reading comprehension)
Self learning feedback loops

My role

Tech lead/CTO, responsible for the overall technical direction of the platform.

Development methodology
Technical architecture
Technology and vendor selection
Hiring of internal, contract and external resources
Partnerships (e.g. AWS & Oracle)

Major challenges

I’ve worked for huge organisations (e.g. EON & Central government) and also Series A+ startups but this was the first time I worked for a really early stage company. I’m used to doing things “the right way” (slow & expensive) but this luxury doesn’t exist at the very early stages of a business. This is especially true when you don’t have a great deal of user feedback and data. Some developers are natural “hackers”, comfortable throwing stuff together at a lightning pace. I’m not one of those people, adapting has been a challenge.

The solution

Viko is a text mining platform that allows users to ask natural language questions e.g. “what is the bin collection schedule?” or “what is the customer service phone number?”. Viko reads through a document corpus, typically website CMS pages, FAQs, Word & PDF documents. It uses these data sources as context, when answering the question.

Viko supports two forms of question answering:

Extractive answers - plucking a fragment of text from a document to answer a question. e.g. given the question “do you offer next day delivery” the answer might be “we offer next day delivery” (assuming this fragment exists in a document)

Generative answers - generating an answer to a question, but using source documents only for context. Using the previous example, a generative answer to the question “do you offer next day delivery?” might be “yes”, although the source document (context) only contained the fragment “we offer next day delivery”

Viko also offers semantic search capabilities, enhancing full text search with reading comprehension. Viko uses machine learning and feedback loops to improve search accuracy. In that sense, Viko is self learning and requires little ongoing tuning

The backend is 80% Python / 20% Scala, with NodeJs (Express) used for APIs and orchestration:

Python
BERT & T5
Tensorflow/Keras
Nodejs (Express/Typescript)
Svelte (Sveltekit/Typescript)

What did I learn?

We made a decision to pivot from a chatbot offering to our current search & question answering platform. The decision to pivot wasn’t based on demand (or lack thereof!) but on supply side constraints. We discovered that onboarding clients and potential clients was just too painful, for all involved. There was too much systems integration, client data feeds were typically poor quality and the feedback loop was too slow.

We got into a vicious cycle - onboaring was slow and costly, therefore only cost effective for large organisations. Large organisations are often slow to respond, making the feedback look even slower. We discovered that the real value of our chatbot platform lay in it’s ability to answer questions. We fell into question answering. We also dicovered that as we’re working with unstructured data, systems integration and data cleansing is much less of an issue.

So that was the big lesson learned - In the early stages of a project, a short feedback loop is essential.