Creating Flamel: A Journey of AI and Design Collaboration
How we merged diverse skills to evolve employee knowledge sharing in our company
In today’s digital landscape, the excitement surrounding AI is palpable. Many are exploring how this technology can enhance their work, while industry professionals are eager to dive deeper, seeking knowledge that can enrich their expertise. Flamel was born from this desire — a blend of curiosity and the drive to experiment.
Everything started when a team of Data Scientists (AI Evolution Hub from Alkemy) reached out to a group of Digital Product Designers (from Design Group Italia, part of Alkemy group):
“Hello! What you do seems interesting: can we do something together?”
Initially, we didn’t fully understand each other’s roles, and, naturally, we spoke different languages. However, we shared a common curiosity to merge our skills and create something tangible. More importantly, we were primarily interested in the opportunity for mutual learning.
Identifying a real need in the Company
We assembled a mixed team of eight — Data Scientists and UX/UI Designers — and launched an internal research and development project aimed at merging AI with design. The concept for Flamel didn’t emerge overnight; we were determined not to rush into the first idea but to address a real problem within the company.
Utilizing the toolkit of a good designer, we conducted interviews with various individuals across different roles (primarily operational) and teams in Alkemy. Our goal was to understand their daily operations, tools, usage of data, challenges, and frustrations.
Through our insights, one prevalent issue resonated across all teams: the need for easy access to internal knowledge. Facilitating knowledge sharing would enhance individuals’ authority through tangible resources derived from past experiences. Currently, each team had created its own systems for knowledge sharing. Given staff turnover, the most valuable resource often became simply asking a colleague for help.
Moreover, analyzing the company intranet analytics revealed that employees required easy access to general information such as news, events, benefits, policies, and more.
Shaping “Flamel”: an integrated AI-based tool
We envisioned a tool that would be easily accessible to everyone. Employees already navigate a complex ecosystem of tools — company chat systems, email clients, payroll systems, time-tracking software, and countless other applications — so it was crucial that our new solution integrated seamlessly into this existing framework rather than adding another layer of complexity.
Ultimately, our idea was to:
Create a virtual colleague accessible through existing chat systems within the company. Leveraging GPT-4 technology, this virtual assistant would tap into a knowledge base of internal content to generate responses.
This knowledge base, called MyLake, was already established within the company and serves as a primary repository, currently housing around 4,000 files — essentially an archive of historical company information accumulated over the years. But more on that later.
Our Minimum Viable Product (MVP) approach allowed us to define an initial set of features through collaborative workshops and prioritization sessions. This process marked our first steps as a multidisciplinary team, enriching each other’s perspectives and sometimes cracking different terminologies. Visual tools greatly boosted our collaboration, also remotely.
How Flamel works: a knowledge base and the retrieval strategy
From our initial research, we identified key motivations for users consulting Flamel:
- Searching for specific documents: Users may seek deliverables or case histories related to specific topics. For example: “Give me the credentials for project XYZ,” or “Can you provide me with offers on healthcare projects?”
- Summarizing accumulated knowledge: Users might want insights derived from historical data. For instance: “How did Alkemy innovate in the healthcare industry?” or “What is the average e-commerce receipt for customers?”
- Accessing up-to-date company Information: Users require current information on company dynamics such as: “What are Alkemy’s conventions and benefits?” or “Who is the BX director?”
But how can Flamel access the correct pieces of knowledge and respond accordingly? This is made possible through a process called Retrieval Augmented Generation (RAG).
While Flamel is powered by the GPT-4 Large Language Model (LLM), it requires more than just that. If you ask ChatGPT any question about yourself or a private document without context, it may respond with something like “I’m sorry but I don’t have information about xxx.”
This is because AI models rely on what they’ve been trained on, which is limited. In contrast, Flamel augments user queries with the most suitable Alkemy documents to provide useful answers. This is called Augmented Generation.
Given thousands of documents in our knowledge base, how do we retrieve the best document(s) to augment user queries? We utilized vector search techniques. Plain text can be transformed into numerical vectors (embeddings) that summarize the semantic meaning of the original text. Similar texts will be represented by vectors that are close together; different texts will be far apart.
When a user asks something to Flamel, their query is first embedded using the same model used for embedding the knowledge base. The generated embedding searches among all document embeddings to retrieve those closest to the query’s vector.
Once we implemented RAG, we needed to define how Flamel would recognize user needs and formulate responses accordingly:
- Retrieve X documents and provide separate summaries with links;
- Retrieve X documents and offer an overview based on them;
- Retrieve X documents and present an overview along with direct links.
After implementing these strategies, Flamel effectively interacted with us. However, we soon encountered another challenge: Flamel didn’t speak our “company language.” Every organization has its own terminology and specific meanings attached to terms or names of teams and roles. For instance, “consulting” isn’t just a generic term for us; it also refers to an internal team.
To address this issue, we needed to tune Flamel to recognize our unique glossary and its meanings. We developed resources for Flamel’s knowledge base so it could learn to communicate in Alkemy’s language effectively.
Establishing Flamel’s identity
The identity of a chatbot as a virtual assistant is crucial for fostering user adoption and trust. From the early stages of defining our chatbot’s identity, we consciously moved away from the traditional robotic representation. Instead, we focused on creating a mascot that embodies trust and security, characterized by a friendly and welcoming personality.
We developed Flamel’s identity around two key ideas:
- The “colleague who always knows how to respond”, because we know from the initial research that people prefer seeking information from trusted colleagues for quick answers.
- The world of Alchemy: to connect with Alkemy’s essence; one name proposal was “Flamel,” referencing Nicholas Flamel — the alchemist who invented the Philosopher’s Stone — symbolizing transformation from raw data into valuable insights.
After settling on “Flamel,” we designed a logo reflecting this personality. We focused on imagery related to flames while humanizing our mascot character to foster emotional connections with users.
Next came establishing Flamel’s tone of voice: spontaneous and informal — aimed at creating smooth conversations akin to those with trusted colleagues. Our goal is for users to feel comfortable asking questions without formalities while maintaining professionalism through clear communication free from technical jargon.
Time for the Pilot phase
We’ve reached a critical juncture: our Flamel MVP is ready for testing! After an intense internal testing phase, we planned a pilot program involving ten employees who would benefit most from using Flamel daily.
The pilot lasted one month during which employees could interact with Flamel in real-world scenarios related to their projects and daily needs. Participants provided direct feedback on their responses while we monitored quantitative results through our Evaluation Kit Dashboard.
At the conclusion of this phase, we integrated both quantitative and qualitative data collected from surveys and individual interviews.
So, how was the feedback?
The evaluation of the overall experience was neutral. On one hand, many users found useful information regarding past projects and appreciated Flamel’s ability to retrieve specific data. On the other hand, some noted that responses to complex questions were not always satisfactory and the initial knowledge base proved insufficient for meeting team requests. Also, privacy concerns emerged regarding data retrieval.
Another important insight is that much operational activity currently resides in subsequent phases where content is processed (e.g., creating slides or case histories).
A plan to evolve Flamel
Our pilot phase has provided invaluable insights, allowing us to outline an evolution program that will enhance Flamel’s maturity. The key areas identified for improvement are:
- Expanding the Knowledge Base — We will include additional corporate data sources such as team directories, project directories, and individual user directories. Each information level will have differentiated access rights to ensure privacy and promote thoughtful sharing of resources.
- Enhancing Data Retrieval Capabilities — We aim to improve data retrieval functionalities by enabling follow-up questions and enhancing Flamel’s performance during uncertainty. Additionally, we will introduce image-to-text capabilities and support additional formats and sources like Miro, Figma, ClickUp, and more.
- Introducing Content Generation Features — We will implement content generation functionalities to facilitate operational tasks, including building case histories and recurring deliverables. Users will also be able to create PowerPoint presentations using the company’s templates and content derived from Flamel’s previous responses.
By focusing on these areas, we believe Flamel can evolve into a more robust tool that effectively meets the diverse needs of our organization while enhancing user experience and efficiency.
Lessons learned & conclusion
Flamel is still a journey in progress; however, as you may have gathered from reading this article, we have already learned numerous valuable lessons:
- It is crucial to look around and identify genuine problems that need solving. In our case, this was also a rewarding human experience — taking a moment to empathize with our colleagues to understand how we could assist them while enjoying the exploration of new technology.
- For this specific product, it has proven essential to comprehend the unique needs of each target team, as everyone operates differently based on their operational requirements. The team must think about creating a scalable and adaptive product, which necessitates an iterative approach centered on user feedback.
- Having a multidisciplinary team has been vital, enriching the experience for everyone involved. Throughout this journey, we leveraged visual communication tools (like Miro) and established a common glossary to align our understanding and enhance effective communication. Above all, maintaining an open mind and a strong desire to experiment has been key.
This article is written by Elisa Fabbian, Giulia Rutigliano and Davide Giannuzzi.
Flamel is the result of collaboration among Ekim Öztürk, Ilaria Porto, Elisa Fabbian, Giulia Rutigliano, Antonio Cappucci (Digital Product Team) and Davide Posillipo, Marcello Villa, Davide Giannuzzi, Milica Cvjeticanin, Stjepan Perkovic, Christian Quaggio (AI Evolution Hub)
It is the product of a research and development initiative by Alkemy and Design Group Italia.
If you’re curious about learning more or believe Flamel could benefit your organization as well, feel free to start a conversation with us at flamel@alkemy.com.
