Making Government Procurement Simple — AI for Contracting
By Emily Hunt Neilson, Director of Government Services
We estimate that our AI-powered models will result in 4x increase in efficiency and annual savings of $52,000 per employee. That’s $26 million saved in 5 years for every 100 employees who adopt the tool.
The acquisition lifecycle in the federal government is notoriously slow and inefficient. In most cases, a relatively simple procurement can invoke a multi-year process. In fact, the government spends more than $500 billion annually — or 40% of all discretionary spending — on acquiring goods and services from vendors.
One common problem that adds to the inefficiency is that federal contracting officers are required to determine fair and reasonable prices for goods and services. They make estimates of the work to be performed by referencing common labor and wage data through manual keyword searches that return an overwhelming number of results then qualify relevant matches.
This is a huge and expensive time suck.
A data-driven AI solution to uncover new efficiencies might be exactly what we need to cut down on such a massive inefficiency, ultimately saving taxpayer dollars and solving challenges in federal government procurement a whole lot faster.
To explore this, RS21 and Google partnered and competed in the Pilot IRS Data Mashing Solution Requirement.
Our team devised a data mashing solution that modernizes the collection, analysis, availability, and visualization of labor and wage data to support better decision-making and lower costs in government procurement and contracting.
To streamline the process, we developed a semantic search engine — an AI-powered prototype using neural networks and natural language processing (NLP) — to improve access and analysis of various sets of labor and wage data.
Unlike keyword and phrase searches that return literal matches, a semantic search generates results by understanding contextual meaning and the searcher’s intent. Take the word “developer”, which could refer to a software developer or perhaps a real estate developer. While a keyword search returns billions of documents containing any instance of the word, a semantic search automatically filters out irrelevant results that don’t fit the context and brings in other relevant terms, such as “software engineer”, in this example.
The semantic search, therefore, allows contracting officers to quickly find and qualify matches so they can focus more of their time on more critical thinking and creative tasks.
Trusted and Explainable AI for Transparent Results
To ensure automated results are reliable, it’s important that users understand the technology and how information is generated so they can defend decisions to peers and decision makers. This requires taking the model out of the black box and developing a defensible and explainable AI model.
One of the best ways to accomplish this is with clear user interfaces and an intuitive user experience. People can then drill down into levels of data to understand how the model is making the prediction and to question it when it doesn’t make sense.
To demonstrate that algorithms are working properly, we look at neural networks and natural language processing (NLP) as layers. Each layer has different types of representation that capture and abstract from the previous layers. When training NLP, we eventually get to word embeddings, which are higher levels of representation that start to capture the semantics of word relationships.
An example is illustrated below where the model begins to recognize the relationship between “king” and “man” and “queen” and “woman”. Similarly, the model shows that it is properly processing geographies, like “Spain” and “Madrid”, and verb tenses, like “walked” and “walking”.
For the Pilot IRS Data Mashing Solution, after testing various methods, we eventually began using flair embeddings. This framework for NLP stacks word embeddings and document embeddings, which performed really well with labor and wage data.
By showing how the model was generating results, contracting officers and other users can feel confident in using the tool to help inform decisions.
AI for Government Procurement
The solution from RS21 and Google is scalable, secure, cost effective, transparent, and relevant for any procurement division. Because the solution automates analysis and minimizes manual input from users, we estimate that our AI-powered models will result in 4x increase in efficiency and annual savings of $52,000 per employee. Assuming 100 employees adopt the tool, it would equate to $26M saved in 5 years.
Moreover, due to the flexible framework and machine learning monitoring and production, the model’s analytic rigor is also expected to incrementally improve over time.
To learn more about our prototype and how it might benefit your agency, connect with me on LinkedIn.
I’d be happy to walk you through a live demo.
About RS21
RS21 is a rapidly growing data science company that uses artificial intelligence, user experience design, data engineering, and modern software development methods to tackle humanity’s biggest challenges. Blending domain knowledge, advanced computational capabilities, and visualization expertise, we develop intuitive and interactive analytics products that provide actionable insights to the Federal Government.
For more information, visit www.rs21.io.
