At Startup Bakery, artificial intelligence is much more than a simple ingredient of the business recipe. We go beyond using AI. We build it and put it in the service of our startups to drive innovation.

So you build Deep Tech startups? No! But we believe AI is an enabling element, a lever to exploit in order to amplify the impact of our SaaS products (on this, we have already written in the past).

And what do you do?

We do AI. To explain exactly and practically what we do, there is no better way than reading QuSeed’s recipe and looking at the ingredients used.

QuSeed is Startup Bakery’s proprietary SaaS supporting us in researching innovation trends. It is in constant evolution, and this allows us to explore cutting-edge technologies, test them quickly and integrate them into the development cycle, then making the new components available across the whole group and thus creating a ready-to-use AI.

Since our SaaS takes in and analyses a large quantity of unstructured data every day (that is, articles detected online), we decided to bring in a set of NLP (Natural Language Processing) techniques to extract value from this information. To this data we also pair financial data, which is in turn manipulated and exploited to identify significant trends in the world of innovation.

Here, then, are some of the main technologies and techniques we adopt for developing QuSeed.

Keyword extraction

Extracting keywords from texts speeds up the consultation and processing of documents, briefly providing the pivotal terms a text is built around. One of the best natural language processing libraries designed to simplify the extraction of meaningful keywords from a text is keyBERT which, building on embedding techniques, uses vector representations of words to rapidly identify a document’s semantic keys.

keyBERT allows us to extract the essential information from the texts taken into QuSeed, simplifying text analysis and reducing the complexity of NLP operations.

Vector databases

Vector databases are a type of database implementing a set of techniques to organise, store and rapidly retrieve information based on vectors, and not only on identifiers or text labels.

Thanks to the explosion of applications using LLMs, generative AI and semantic search, vector databases have become a must for efficiently processing the volume of data generated by models and for enabling the introduction of advanced features such as semantic information retrieval, long-term memory, clustering and so on.

Inside QuSeed we use vector DBs to store the embeddings of the articles to be analysed, and we exploit them for content search (semantic search) and to provide context to our chatbot (Austin), making it able to answer questions concerning that content.

Semantic search is an approach to information retrieval that aims to understand the meaning and context of a search query, in order to return results that are more pertinent and meaningful than a search based on keywords alone. This technology seeks to interpret the meaning of the words in the query and to provide results more coherent with the user’s real intent.

QuSeed’s free searches use precisely this approach, making it possible to give the user pertinent answers, both when searching financial data and when navigating the news we take in.

OpenAI API and LangChain

OpenAI’s APIs offer access to powerful language models with natural language processing (NLP) capabilities. These models are able to respond intelligently to natural-language inputs.

LangChain is a framework designed to simplify the creation of NLP applications using Large Language Models (LLMs) that are context-aware (they give answers suited to the context) and cross-cutting (they allow the use cases of LLMs to be broadened).

Both, combined with QuSeed’s robust internal database, power our proprietary chatbot Austin.

OpenAI allows Austin to understand and respond naturally to human inputs, creating a fluid, intuitive communication experience, and allows users to interact with the chatbot as they would with a human being, without having to adapt their language or terminology.

The LangChain framework makes it possible to manage and chain the requests sent to OpenAI, ensuring every conversation with Austin is coherent and well contextualised.

Statistical analysis and time series study

Statistical analysis is the method that allows data to be collected, analysed and interpreted. One of its branches is the study of time series, which deals with data that changes over time. These techniques make it possible to identify patterns and correlations in the data, in order to better understand the phenomena underlying it. In the startup field, they can be used to identify trends, new investment opportunities, emerging sectors and so on.

QuSeed uses statistical methods such as the Mann-Kendall test to perform an in-depth analysis of time series, detect patterns and anomalies and identify the presence of hidden trends in the data or significant events over time. It also uses time series derived from proprietary metrics tied to the articles analysed, measuring aspects such as the Information Diffusion Rate or user interest.

QuSeed also carries out a detailed study of the correlation between different time series, providing precious insights on the relationships and interconnections between the various data points and thus giving a complete, in-depth view of the information landscape covered.

SetFit and Contrastive Learning

SetFit is a powerful, versatile framework that can be used to train a variety of machine learning models on small datasets. Contrastive learning, meanwhile, is a machine learning approach based on differentiating between similar and different examples, with the goal of learning to distinguish between examples belonging to the same category and examples belonging to different categories.

Both are used in the learning process of Bouncer, the proprietary machine learning model within QuSeed for advanced filtering of all the document data acquired, making it possible to have enough examples to reach high precision despite a starting dataset of limited size.

Bouncer’s main goal is selectively eliminating the data deemed not useful for our analyses, ensuring only the most relevant information is included in our studies and offering a targeted, optimised approach to managing information.

Clustering

Clustering techniques are methods used in data analysis to group sets of observations that are similar to each other. We use a hybrid approach to clustering, which includes DBSCAN, a technique identifying clusters on the basis of the density of the elements, combined with custom algorithms to dynamically manage the addition of documents to the identified clusters.

Clustering makes it possible to identify QuSeed’s topics, that is, the sets of articles from which the trend analyses start.

Research and development are the foundation of our work

All these technologies and techniques allow QuSeed to:

extract the essential information from the texts taken in

store and retrieve information efficiently

search for information in a pertinent, meaningful way

understand and respond naturally to human inputs

identify trends, new investment opportunities and emerging sectors

selectively eliminate data deemed not useful.

QuSeed is a tool that can be used by startups, companies and investors to identify innovation trends and make informed decisions.

That said, the technologies listed in this article are only some of those we adopt and that make up both QuSeed and the AI components made available to each of our startups.

The choice of every technology, framework and technique is the fruit of months of research and testing, aimed at identifying the best road in terms of performance, integrability, quality and costs.

(Sustainable) technological innovation is the main ingredient at Startup Bakery, and precisely for this reason we invest constantly in research and development, with the goal of identifying the most advanced technologies and techniques to generate the best business ideas and support the startups of our startup studio on their growth path.

Startup Bakery is the Italian startup studio specialised in creating B2B SaaS companies with Artificial Intelligence. We offer aspiring Co-Founders the opportunity to develop a business idea. We create investment opportunities for Professional Investors. We help companies in their innovation process.