FASTer

A Usable, Accessible and... FASTER UI Experience! And yes... it’s AI powered

Comment

Inspiration

Internal tools should be improving efficiency and reducing friction, not causing it. No one enjoys doing busy work and even fewer people enjoy doing it on a clunky outdated User Interface. That is why we set out to Research, Plan, and Develop FASTer, a full UI/UX refactoring of First American's FAST internal tool.

What it does

Our redesign delivers a FASTer, Safe, Usable, AI accelerated User Experience! It accomplishes this via the following:

  • Our newly refactored user experience and site layout. This reduces the headache of reading a screen length bullet point list of menu options. Instead, we utilize a header navigation menu to make the website usable, in modern standards.
  • Our Parsing Agent, handles the parsing and cross validation of POAs and their subsequently generated Deeds for Escrow Agents working on the convoluded Real Estate transaction process. They no longer have to manually parse Power of Attorney (POA) documents or manually validate generated Deed agreements.
  • We also provide Escrow Agents with a Risk Analysis Agent that conversationally allows them to ask information relating to the documents involved with any given transaction and be returned a risk score. These Agents, speed up the real estate transcation while reducing the likelihood of user error.

How we built it

  • Beginning with research we first needed a solid understanding of the Title Insurance Domain, the documents it involves, and of the Real Estate transaction process. To achieve this we conducted research online, through Gemini and First American Employees.
  • To understand the tool we were redesigning we focused on its core functionality, inputs, and outputs when surveying current First American employees. We then utilized Figma Make to quickly generate a starting point for our redesign and Gemini CLI to customize the shell design.
  • To integrate AI and set apart our design we utilized our stakeholder conversions to derive a set of pain points and flagship features described above. We planned, architected and generated our API Agent Backend through the usage of AI coding tools like Claude Code and Codex, ensuring to utilize proper and responsible AI usage.
  • To address our lack of data we first sourced official First American POA and Deed document templates. We then utilized Claude Code to generate script that returns a set of sample transaction scenarios and its accompanying documents. Leaning on our mock data/scenarios we were able to complete a functional MVP (Minimum Viable Product) that showcases everything described above.

Challenges we ran into

  • We as a team had no previous knowledge or understanding of the title insurance domain. To address this, we conversated with LLMs like Gemini and ChatGPT to source a rough understanding. For intricacies, we conversated with First American Employees and learned from the source.
  • Furthermore, due to the nature of the domain there were no official statewide databases of California deeds/related documents for us to utilize. As described above, we generated mock data that aligned with First Americans document templates to address our lack of data.
  • Our final problem was lack of access and information of the system we were redesigning. Again, through our resrouceful research methods, we were able to gather first hand information and ultimately understand the existing FAST system.

What we learned/Reinforced

From this Hackathon our takeaways are categorized as follows:

Technical - Python Libraries: FastAPI, Pdfplumber, Anthropic SDK, Python-dotenv, Pydantic, Uvicorn

  • We built our entire backend API using FastAPI, learning how to handle async background tasks, multipart file uploads, and polling endpoints to support long-running AI analysis jobs without blocking the server. - We used pdfplumber to extract raw text from uploaded property PDFs, handling edge cases like image only documents and multi-page files.
  • We integrated the Anthropic SDK directly to call Claude Haiku, learning how to craft structured system prompts that constrain the model to output valid JSON and ground its responses in a curated knowledge base rather than general knowledge.
  • Pydantic was used for request validation, ensuring all incoming API data was typed and safe before hitting our business logic.
  • Python-dotenv taught us best practices for secrets management — keeping API keys out of source code entirely.
  • Uvicorn served as our ASGI server, and we learned how to run it in reload mode during development for fast iteration.

AI Agents

  • We learned how to design a RAG-grounded agent — rather than letting the model reason freely, we injected a curated knowledge base of 225 known property risk patterns directly into each prompt, forcing the agent to only flag issues it could cite by ID. This made the output auditable and defensible.
  • We built an asynchronous agent pipeline where document ingestion, risk scoring, and summary generation all happen in a background task, with the frontend polling for results teaching us how to architect agents for real-world latency constraints.
  • We implemented a multi-turn conversational agent grounded in both the original document text and the completed analysis, so follow-up questions stay contextually accurate rather than hallucinating answers.
  • We learned the difference between sync vs. async agent patterns — using background polling for the heavy analysis pipeline and a synchronous endpoint for the real-time chatbot, and when each approach is appropriate.
  • We applied prompt engineering techniques like output format enforcement, constraint injection, and role-based system prompts to make the agent behave predictably enough to ship in a demo environment

Software Design Life Cycle

  • Requirement Elicitation
  • Proper and Effective Stakeholder Communication
  • Effective project scoping
  • UI/UX research
  • Deliverable iteration

Accomplishments that we're proud of

We as a team are proud of our ability to architect, and ultimately deliver a project as ambitious ours. We did not shy away from the ambiguities, rather we tackled them head on with innovative and modern solutions. In under a weekend we were able to:

  • Learn about the Title Insurance Domain and the Real Estate transaction process
  • Learn and recreate the functionalities of the existing FAST system
  • Create a working UI and subsequently overhaul it
  • Integrate AI in a creative and innovative way And finally, provide a working proof of concept to highlight the impact AI technologies like Agents and LLMs can have in the domain of Real Estate or Title Insurance!

What's next for FASTer

To effectively demonstrate our vision within the hackathon timeframe, we built our demo based on mock data and controlled scenarios. We designed the architecture under the assumption that a future phase would involve connecting our frontend directly to First American's existing FAST backend. If given the opportunity to fully realize this project, our immediate next step would be to replace the mock scenarios by integrating the existing FAST database. This would improve our AI agents' output quality because they would now have access to data that could further their thinking abilities.

It is also important to note that a redesign is only successful if it actively accompishes its goal, improving the user's experience. Therefore, our next step for our UI/UX redesign is to put our initial build into the hands of our primary users. We plan to conduct structured user testing to gather direct feedback on our new workflows, allowing us to refine and finalize the interface through multiple rounds of iterations.

Built With

Share this project:

Updates