GovCon AI

Inspiration

One of our teammates' family member works in the government contract space in a local small company. However, the company has been struggling to find and submit suitable requests for proposal, and this is an existential threat to their business, exacerbated by massive firms competing in the space. We built this tool to level the playing field for smaller companies in this space.

What it does

The platform starts by building the firm "DNA" by analyzing the firm's description, employee resumes, and past proposals that had been submitted. Then, a web scraper scrapes RFPs (requests for proposal) from sem.gov and displays the ones that are deemed as most relevant and useful for the firm. The built-in RAG system helps draft the RFP application, which is done by extracting requirements, building a compliance matrix, and citing employees for technical sections. The system is designed so that it won't hallucinate because it only pulls data from the firm DNA.

How we built it

We developed all components locally then deployed to AWS using cloud development kit (infrastructure as code). Frontend: Next.js + React Infrastructure: SST on AWS AI: AWS Bedrock (Claude) Database: Aurora Postgres Writer: ECS Fargate Auth: Clerk

Challenges we ran into

Initially, we wanted to web scrape every single RFPs in sam.gov. However, this started to become unfeasible because of the scale of the number of RFPs on the website and realized that there would be no real benefit for companies. As a result, we decided to change the algorithm so that it would utilize the client profile and develop a RAG system that would detect the most relevant RFPs and only scrape and display those on our dashboard.

Accomplishments that we're proud of

We scraped thousands of RFPs form sam.gov site, developed a high quality RFP drafting system based on compliance matrix derived from RFP requirements, and deployed multi-agent system for quality assurance and pricing estimation.

What we learned

We learned that agents perform best when their tasks are narrow, requirements are clear, and source data is robust. We quickly found that trying to brute-force data acquisition through web scraping without an intelligence layer was not feasible, this motivated us to develop a more intelligent system that could act more efficiently based on the specific user's firm capabilities. We found that qualitative information can help for generative tasks like document drafting, but a quantitative set of requirements with specific numbers and binary requirements (the compliance matrix we developed) is essential in yielding valuable results and mitigating hallucination.

What's next for GovCon AI

We plan to extend to platform to include tools that work in word processing and spreadsheet applications to decrease friction of using our platform and increase value by serving as a drop-in solution for firms that are already spread thin and are weary of adopting new tools that require intensive training.

Built With

• ai:
• aurora
• bedrock
• cdk
• clerk
• ecs
• fargate
• lambda
• next.js
• postgresql
• react
• sst