Here at Metalab, the success of our work, and ultimately our partners' products, comes down to our ability to assign the right people to the right projects at the right time.

Sounds simple enough, but all our clients have very different needs, requirements, and requests.

At any given time, we can have 15 to 20+ different types of projects happening with different start dates, phases, and completion dates. 

Add our team’s vacation, holidays, and sick time to the mix, and it gets quite complicated.

To manage all of this, we use several different tools and third-party systems:

Sales and CRM platform to manage opportunities and clients

Google Sheets for financials and project estimations

Secure document management for contracts and client requirements

Slack channels for sales, leadership, and resourcing teams to discuss each opportunity

If you work in an agency or consulting firm, this might sound familiar.

Our resourcing team spends many hours a day analyzing 20+ Slack channels to understand each project's needs, role allocations, and client requirements.

Then there’s the many cross-department resourcing meetings.

It’s manual, tedious, and time-consuming—the perfect challenge for AI to solve.

Our Approach to Building a Resourcing AI Assistant Slack App

Our teams primarily engage in detailed project discussions and requirement analysis on Slack. To avoid introducing additional overhead with another web application, we developed our AI Assistant as a Slack app, aiming for seamless integration.

We leveraged ChatGPT by creating a resourcing agent aimed at producing informed resource allocation reports. We introduced the concept called “Retrieval Augmented Generation” (RAG) that enabled us to supercharge our large language model (LLM) with our business internal project and people data.

We broke things down into five key steps:

Step 1: Data Collection — Laying the Foundation for Insightful Analysis

Step 2: Data Preprocessing — Harnessing the Right Technology for Efficient Data Transformation

Step 3: Feature Engineering — Understanding the Problem for the Correct Solution

Step 4: Model Selection — Optimizing Data for Maximum Predictive Power

Step 5: Training — Constantly Improving and Enhancing The Results

The initial step is always data collection. Data is crucial for artificial intelligence and machine learning, as historical data enhances machine learning predictions and improves the model’s ability to generate insights.

These are the systems that we needed to integrate:

Airtable: People skills and previous experience data

Slack channels for each pursuit: Long conversations between pitch team members that typically had the context of client needs that spanned weeks

Google Drive to PDF and Docs of Statement of Work, RFP, requirements

You can get pretty good results and insights out of the box by plugging your data into any Machine Learning, AI, or Statistical model. The problem is that data barely comes in the correct format and is not always on defined cycles.

Most of the time, it will be messy, missing, or just plain incorrect. In other words, 90% of the time, you will have to work with unhappy families, meaning that you will have lots of unprocessed data that comes in different formats from different sources.

In our case, we had to extract and transform data from PDF files and process information in the form of messages in Slack, resourcing tool APIs, and other third-party channels.

This meant transforming raw data from all our sources into skills, capabilities, roles, and activities that the model could use to process.

We explored various cloud-provided tools in AWS, such as Kinesis, Firehose, Analytics, EMR, Glue, and Athena, to efficiently collect, collate, and query data. Understanding these powerful tools enabled us to utilize the data effectively. 

Given the large amounts of textual data, we chose an LLM as our base model.

Feature engineering is a critical step in the machine learning pipeline, where domain knowledge is used to create, select, and transform raw data into meaningful features. This process enhances model accuracy by providing algorithms with more relevant information to make predictions.

Effective feature engineering involves understanding both the data and the model's specific requirements. This ensures that the input data is optimally structured to highlight underlying patterns related to the predictions.

For us, this required extracting and transforming variables from raw data, such as skills, capabilities, roles, activities, and availability, so that features could be used for training and prediction.

Once we gathered and processed all the information in the correct way, we had to select the model.

We realized that most of the information was in huge amounts of text paragraphs, so it made sense to use large language models (LLMs). These are models that have been trained with a lot of information in the past and can understand the context of whatever text you pass through them.

After selecting models, we enhanced copy capabilities using 

, a library that automates the process and helps the LLM progressively digest the information. We also used other frameworks, like 

 to retrieve the information quickly and efficiently.

In the end, we were able to create regressions to make predictions about the future. We used transformers to transform the text information into numerical data that is much easier for the model to understand. Then, we used clustering to identify different types of text that eventually become features, skills, or characteristics of the project.

These features allow us to compare projects and aggregate information in a single place to return useful answers.

The final step is the training process. No matter the problem that you’re facing, you are always going to be dealing with this phase.

Once you collect all your data, train and select the model, and apply the training process, you’ll get some kind of result. The first time that you apply your model, you’re not going to get what you are expecting.

However, it will give you good insights into what path you have to follow next. This means that every time you get new results, you need to uncover what the model is trying to tell you, which features or variables are useful for the model, and what information you need to create better predictions.

At this point, we started working closely with our resourcing team to gather better features and understand what information was available. This helped us understand what problems we were trying to solve.

This is the interactive process of training model output and evaluation.

, an open-source repository for models trained on thousands and thousands of files. We had to convert all that information from contract documents, Slack channels, and Airtable into a format that was comparable between different documents or between different texts.

We obtained two different representations: one is a service that allows us to answer a specific question, and the other is for SOW documents that allow us to include the deliverables, activities, and goals.

Then, we started an iterative process to acquire more information and compare it to the historical data. This allows us to leverage previous project history to enhance the results every time we retrieve information.

We believe in iterative, constant releases so we can continually refine and learn. Our product started with two key features to begin with:

Weekly Resourcing Digest integrated with Slack

Generate Resource Allocation for each project opportunity

The first utility feature is called the “Weekly Resourcing Digest” and integrates directly with Slack.

On Fridays, the platform checks all of the projects that we currently have open and all of the Personal Time Off (PTO) that we currently have. It analyzes who will be coming off the bench or off a project and who will be available for assignment on a new project.

Our leadership and sales team in the appropriate Slack channels get a weekly digest of who’s going to be coming off of projects and who is going to be available. This utility feature helps them understand who’s coming available and how they can start thinking about resourcing for the next week.

The next key feature is called the “Generate Resource Allocation” and is added to each pursuit Slack channel.

In each of these Slack channels, the assigned pursuit team typically discusses the requirements and needs of the client. Once the team negotiates with a client to align on the statement of work contract (SOW), they are ready to assign the right project team to deliver the work.

This screen pop-up in Slack uncovers information about the project specifics to make a representation of that project.

You can see here that the LLM has parsed all of the uploaded information from the Slack channel to pick the most relevant options for these fields to represent the project, i.e., “Al/Machine Learning, Healthcare & Wellness, Mobile Apps and Web Apps.”

Once that is done, the pursuit team can add anything that they see is missing. Then, they enter the tentative start date.

This report generates in the background and is what we’re calling the retrieval augmented generation (RAG). That is how we’re supercharging the LLM.

The platform is making API calls to BambooHR for PTO or to our Amazon S3 Data Lake to capture all relevant information and ensure that our resourcing specialist (the AI) has enough data to make accurate predictions.

If the AI didn’t know who is available, who’s off, who’s on a project already, and what their skills and requirements were, it would be harder for it to make that decision. Retrieval augmented generation fetches the business requirement data first and injects it into the prompt before sending it off to the LLM.

Based on the project needs, the LLM, or the Resourcing Assistant as we call it, has chosen roles that are specifically required for this type of project. Then it went a step further to select people based on their skills, experience, and capabilities in the industry that are relevant to this specific project.

It is suggesting our Front End Engineering Lead, Backend Principal developer, Frontend Principal developer, Senior Backend developer, and Senior Front End developer for the project. It is also making some consideration notes that help the resourcing team understand why these choices were made.

That way, in the future, they can start to realize what the LLM was choosing for these people based on their specific attributes, so they can start to better recall what our people are capable of and map those connections to their capabilities.

Providing all of these people are available, the Resourcing Assistant would make the best selections for that team. It also suggests backup candidates for future consideration. That way, if there are people who might already be staffed on a project and won’t be freed up soon, we can still suggest them as backups in case they become available. It uses the same logic in order to place people based on their skills, their experience, and past projects that they’ve worked on.

The Impact of Our Resourcing AI Assistant

This should give you insight into one of Metalab’s hardest clients — ourselves. Like all of our product work, we are constantly learning, getting feedback, and refining the product to be better.

We measure success the same way—based on the impact it has on the business and its users.

Since shipping our AI Resourcing Assistant, we’ve saved hours every week discussing and performing tedious, redundant staffing and resourcing tasks. This has freed up our time from manual work so we can focus on higher-impact work, like people’s skills, experiences, and areas of growth. 

This allows us to be more strategic for the business when it comes to staffing decisions and helps us to be more nimble in our sales process. Ultimately giving us space to better support our clients and people’s needs.

Stay tuned! We’ll share a few related posts soon on our 

The Metalab Machine Learning Framework in creating great products

How to Build an AI Assistant that Drives Immediate Value