This post was written by Vincent Seguin, Bridgit’s VP of Engineering
Table of Contents
At Bridgit, we’ve been working incredibly hard over the past several months to achieve our first AI launch, and we couldn’t be more excited to share it with the world. It’s an important milestone for our company, but more importantly, for everyone who relies on our software for the important task of planning project teams and winning work.
Rolling out the first artificial intelligence specifically built for construction workforce planning is no small task, and we’ve already heard a lot of questions about how we built it, and what makes it unique. This post will tackle many of those questions from a technical perspective, and will also explain our journey into a company with AI at its core.
Note that this post will primarily focus on Ask Bridgit, our flagship and most complex AI offering. But much of what I’ll discuss is also a reflection of our broader AI strategy.
Just Ask Bridgit
Ask Bridgit is the perfect vehicle to debut Bridgit AI, because it exemplifies so much of what we’ve built our company around. There’s a tremendous amount of valuable workforce data within every construction company, but far too often that data is overlooked or not structured in an accessible way.
Ask Bridgit gives anyone the ability to ask direct workforce-related questions – like who’s available for an upcoming project or who has specific experience – and get a concise answer. It essentially puts everyone on a level playing field, so it’s no longer only those who have been at your company for years that can access this valuable institutional knowledge.
Becoming an AI-first organization
Like many companies with established software tools before the arrival of accessible GPT-style AI models, we identified a necessity for innovation. Embracing AI was non-negotiable, but we couldn’t do so in isolation. We understood that artificial intelligence had to become an essential part of who we are as a company. This meant incorporating AI into the heart of our strategy, operations, and decision-making, and making it a priority to help our people upskill and experiment confidently.
We subsequently launched several initiatives designed to spark curiosity and make learning part of our everyday flow:
- Shared learning channels – We engaged in dedicated Slack channels including #dev-ai, #dev-learning, and #learning-ai, where team members shared ideas and collaborated.
- Expanded learning budgets – We doubled yearly allocations to give everyone more opportunity to expand their knowledge.
- Experimentation funds – We also budgeted to experiment with new AI tools via a fast, secure review process.
- Knowledge sharing – We shared Loom demo videos across the company to showcase experiments and findings.
- Developer enablement – We rolled out Cursor, Junie, and Claude Code across our engineering team to accelerate the adoption of AI-assisted development.
These steps turned curiosity into momentum – and momentum into progress – which helped lay the foundation for what came next.
Building on our data foundation
When building any AI system, the very first question to ask is: do you have the data needed to make it successful? Without strong, reliable data, it’s nearly impossible to create something valuable. Fortunately, we had a head start. At the beginning of 2025, we launched Experience Tracking, a first-of-its-kind way to track dozens of key data points relating to the experience of construction team members. This unique combination of experience and workforce planning data provided a powerful foundation for our AI journey.
On the topic of data, it’s crucial for me to stress that data security is at the very core of everything we do at Bridgit. We don’t train any proprietary models ourselves; instead, we leverage pre-trained, widely available models. Most importantly, no customer data is ever used to train this system – and all of our existing safeguards, including SOC 2 and GDPR compliance, fully apply.
Diving into the details
Let’s get to the nitty-gritty. Our AI foundation is built on Amazon Bedrock, primarily using Anthropic models, which allows us to stay within our existing secure AWS infrastructure. Ask Bridgit itself is built in C#, following the same standards as our other services. This flowchart demonstrates how we handle an Ask Bridgit query.
- Query validation – Each query is first validated to ensure it contains no profanity or malicious intent.
- Retrieval – Using RAG with pgvector and Cohere’s embedding model (embed-english-v3), we retrieve several pre-made example SQL queries that best match the user’s question.
SQL generation – We then leverage Claude Sonnet 4.5 to craft a SQL query, providing the original question, retrieved samples, available SQL functions, and detailed instructions on what to do (and what not to do). - SQL validation – The crafted SQL query is sent back to Sonnet for a validation and optimization pass.
- Execution – Once validated, the query is executed against a read-only replica of our database. The user running these queries has highly restricted permissions, with read-only access and limited to a subset of SQL functions specifically designed for this use case. These functions are scoped by account_id to prevent any cross-account access.
- Post-processing – The results are post-processed to ensure that no data from other accounts has been returned.
- Summarization – Finally, the results are sent to Llama Scout – a very fast model – to produce a quick, human-readable summary.
Simple enough, right? Supporting new questions is straightforward, we just add new examples or extend our set of SQL functions as needed.
What’s coming next?
We couldn’t be more excited about launching Bridgit AI and reaching such a significant milestone as a company. What’s even more exciting is that this is only the beginning. We’re already working to expand the capabilities of Ask Bridgit, so it can have an even greater impact to all of our customers.
We’re also experimenting with a feature we’re tentatively calling Do Bridgit (name TBD!). This exploration is aimed at tackling more complex, agentic workflows that can take action on behalf of our users.
As we look ahead, there are unlimited opportunities to explore and questions to answer. One thing’s for certain: this journey is only getting more exciting with time, and there’s much more to come.
