Home Sweet Home

Chatbot with Bedrock Knowledge Base

🔗 Prototype

✏️ Base Knowledge

What is 청약 (Cheongyak)? Cheongyak is a housing subscription system in South Korea that allows individuals to apply for new housing developments under specific eligibility criteria. It provides opportunities to secure homes at prices lower than market rates, making it a vital resource for many.

💡 Inspiration

applying for Cheongyak is often tedious—checking for new announcements, reviewing over 50 pages of requirements, and verifying eligibility can be overwhelming. Upon researching the market, we found that existing Cheongyak alert services primarily focus on expensive private housing sales. To address this gap, we developed a service tailored to young applicants, offering personalized alerts to simplify the process and ensure users receive only relevant opportunities.

👩🏻‍💻 What it does

Our service streamlines the Cheongyak application process by automating key steps. Every day at a specific time, we crawl bulletin boards from platforms like LH and SH to gather newly posted Cheongyak announcements. These announcements are then matched to users based on their preferences and eligibility criteria. Matched users receive an organized email containing links to the relevant Cheongyak announcements, along with a chatbot link designed to simplify communication. Through the chatbot, users can easily find out the application deadline, required paper works, and other important details—all without having to shift through lengthy documents or complicated processes. This service makes navigating Cheongyak faster, easier, and more accessible for everyone.

🛠️ How we built it

Crawling & Data Storage

EC2: Automatically crawls bulletin boards (e.g., LH, SH) daily to collect Cheongyak announcements, including titles and links.

DynamoDB: Stores crawled announcement metadata.

Amazon S3: Hosts raw PDF files of the announcements.

PDF Processing & Data Extraction

S3 Trigger: Detects new PDF uploads and initiates processing.

Upstage API: Extracts key details (e.g., eligibility criteria, deadlines) from PDFs.

DynamoDB Update: Enriches announcement records with extracted data.

User Profile Management

Frontend: Collects user preferences and eligibility criteria.

User DynamoDB: Stores user profiles for later matching.

Automated Matching & Alerts

EventBridge: Monitors DynamoDB for updated announcements.

Lambda: Matches announcements to eligible users using queries.

Email: Sends personalized alerts with announcement links and a chatbot link.

Chatbot Assistance

Chatbot: Lets users ask questions (e.g., “Do I qualify?” or “What’s the deadline?”) and instantly receive answers based on PDF data.

🔥 Challenges we ran into

PDF Integration & RAG Complexity

Initially, we attempted to directly connect raw PDFs to Amazon Bedrock’s general models for RAG (Retrieval-Augmented Generation). However, inconsistent formatting and extraction errors made this approach unreliable. To resolve this, we pivoted to using Bedrock Knowledge Bases, which streamlined data structuring and improved accuracy.

Model Permissions & Regional Constraints

Certain Bedrock foundation models required explicit IAM permissions or were restricted to specific AWS regions. Navigating these limitations—submitting permission requests, adjusting region configurations, and testing compatibility—added unexpected delays to our development timeline.

User-Announcement Matching Accuracy

Matching user profiles to eligible Cheongyak announcements required precise prompt engineering. Early iterations often produced false positives/negatives due to ambiguous criteria (e.g., income brackets). Through iterative testing, we refined prompts to better interpret eligibility rules and reduce mismatches.

These hurdles ultimately strengthened our system’s reliability, ensuring seamless interactions between users, data, and AI models.

🏆 Accomplishments that we're proud of

Precision Prompt Engineering for Bedrock

By meticulously refining prompts based on user profile attributes (e.g., income level, residency duration, savings history), we drastically improved Bedrock’s ability to match announcements to eligible users. For example, adding descriptions of each key into structured prompts allowed the model to evaluate eligibility with 98% accuracy. This turned a subjective process into an objective, scalable solution.

Seamless Integration Under Tight Deadlines Despite the complexity of our tech stack—spanning EC2, DynamoDB, Bedrock, and Lambda—we successfully unified all components into a cohesive workflow in just 15 hours. The moment our first end-to-end test delivered perfectly matched alerts with chatbot links, we knew we’d achieved something extraordinary. It wasn’t just technical execution; it was teamwork, adaptability, and shared vision in action.

🧠 What we learned

The Critical Role of Permissions

Navigating AWS permissions taught us that resource access controls can make or break a project. From Bedrock model restrictions to IAM role policies, we realized how easily misconfigured permissions can block progress—and why the "principle of least privilege" is both essential and challenging to implement. This experience deepened our respect for AWS security best practices.

⭐️⭐️⭐️Prompt Engineering: The Unsung Hero of LLMs As LLMs like Bedrock become central to tech solutions, we learned that prompt quality directly dictates output reliability.

Embracing Challenges Accelerates Growth Struggling with Lambda layers and container deployments initially felt overwhelming. However, trial-and-error—like rebuilding layers 10+ times or switching to containerized Lambdas—taught us more than any tutorial. Now, we’re confident in deploying serverless functions efficiently.

🔜 What's next for Home Sweet Home

Expanding Beyond Youth Targeting

While currently focused on youth-specific opportunities, we aim to broaden the service to include more general housing subscription programs, making it accessible to a wider audience.

Global Adaptation

housing subscription systems like Cheongyak exist in other countries. We plan to research and adapt our platform to support international housing programs, enabling users worldwide to benefit from tailored alerts and assistance.

Enhanced UI/UX

Improving user experience is a priority. We will refine the interface to ensure it is intuitive, visually appealing, and user-friendly, allowing seamless navigation and interaction for all users.

These steps will help us scale the service while maintaining its core mission of simplifying housing applications.