BalaOne

Inspiration

As aspiring 17-year-old entrepreneurs, we quickly realized that every new founder in Hong Kong hits the exact same wall: you can register a business over a weekend, but finding the right commercial space takes weeks. Hong Kong is home to about 357,000 SMEs, and with entrepreneurship accelerating (over 5,200 startups projected in 2025), a massive wave of first-time tenants and buyers are entering the market. They need to move fast.

Yet, the commercial property search system is painfully outdated, fragmented, and agent-led. The same unit shows up multiple times across platforms, key specs are inconsistent, and crucial operational feasibility questions are only answered after time-wasting viewings and endless back-and-forths. This outdated process wastes founders’ time, delays grand openings, and weakens negotiation leverage before they even have a confident shortlist. We wanted to build an AI co-pilot that turns weeks of guesswork into minutes of confident, decision-ready action.

What it does

An AI advisor for navigating Hong Kong’s commercial property market, turning fragmented listings into a ranked, explainable shortlist and a next‑step action pack tailored to the user’s needs.”

First, a multimodal input is entered with all the user’s specifications, where the AI then asks targeted follow-ups which leads to the generation of a clean shortlist—, ranked, and tailored—so users are not scrolling endlessly through irrelevant options.

When users click a property, the AI provides the key benefits, the trade‑offs, and—most importantly—the unknowns, users should ask the agent before wasting time with a viewing. It also shows a regulatory and legal checklist of what’s confirmed versus missing, and even includes Feng Shui for buyers who care. Users can go on to explore trade‑offs with ‘What if’ scenarios—like increasing budget, or giving up a panoramic view—which updates the shortlist immediately and can be reverted, allowing users to see better options mimicking real-life refinement. Finally, users can compare any two properties side‑by‑side with an 1‑v‑1 verdict and the numbers that matter—generating an analysis of the two selected properties and comprehensive financial advice advising firms to buy, rent and/or what types of mortgages to take out.

How we built it

To build BalaOne rapidly and efficiently, we combined modern design tools, AI-assisted development, and robust cloud infrastructure:

1. Design & Ideation: Before writing a single line of code, we used Canva to design our UI/UX wireframes, craft our branding, and visually map out the user journey.

2. Rapid Development: We built the entire frontend and backend using Cursor, the AI-powered IDE. This allowed us to iterate incredibly fast on our Next.js architecture and rapidly write the complex logic needed for our AI integrations.

3. Database & Authentication: We implemented Supabase as our core database (PostgreSQL) and backend service. It securely handles user authentication, stores our complex property datasets, and manages user sessions, saved shortlists, and property evidencePack records.

4. Multimodal AI & Chat: To handle the dynamic, conversational front-end of our co-pilot, we integrated MiniMax AI models. MiniMax powers the initial multimodal inputs—taking the users' budgets, timelines, and inspiration images—and asks targeted follow-ups to extract hard filters before handing the data off to our stricter analysis engines.

5. Specialized AI Agents (AWS): For the heavy-lifting analysis, we built a custom backend powered by AWS Bedrock:

• Core Client: We set up src/lib/ai/bedrock-client.ts using the BedrockRuntimeClient and ConverseCommand to securely manage credentials and model routing.
• Lease Reviewer: Our POST /api/ai/lease-review endpoint uses Bedrock to run parallel "Legal" and "Compliance" agents. It takes raw lease text and outputs a structured breakdown of legal issues and HK regulatory red flags.
• Risk Engine: Our POST /api/ai/risk-check passes property data (including UBW status and building records) to Bedrock via risk-engine-bedrock.ts, generating a comprehensive fit-for-use pass/fail score with actionable recommendations.

Challenges we ran into

Challenges We Faced

1. Taming Unstructured Data: Hong Kong commercial listings are notoriously messy. Extracting reliable variables (like gross vs. net area, or MTR proximity) to feed into our Supabase database and AI risk engine required us to build robust data-cleaning pipelines.

2. Prompt Engineering for Strict Output: Getting the AWS Bedrock models to consistently return structured, parseable JSON for our Legal and Compliance agents (instead of chatty, conversational text) took dozens of iterations of system prompt tuning.

3. Handling Complex HK Regulations: Hong Kong property law and commercial zoning rules are incredibly dense. We had to ensure our AI didn't hallucinate regulations, especially regarding things like UBW (Unauthorized Building Works), which required us to carefully inject factual context into our AWS Bedrock Converse commands.

Accomplishments that we're proud of

Accomplishments That We Are Proud Of

1. Building an Enterprise-Grade Architecture at 17: We successfully integrated a complex, multi-model AI pipeline (AWS Bedrock and MiniMax) with a robust Supabase backend, proving that age isn't a barrier to building highly technical, scalable solutions.

2. Automating Complex Legal & Compliance Analysis: We are incredibly proud of our proprietary AI Lease Reviewer and Risk Engine. Successfully engineering an AI to accurately parse dense Hong Kong commercial lease text and flag critical Unauthorized Building Works (UBW) risks is a massive technical win.

1. Creating a True "Minutes, Not Weeks" Solution: We took a notoriously fragmented, agent-led process and condensed it into an intuitive UI. Seeing our platform successfully generate ranked, explainable shortlists and financial verdicts out of messy, unstructured listing data validated our entire vision.

What we learned

1. Mastering Cloud Infrastructure: Integrating AWS Bedrock and Supabase was a massive learning curve. We learned how to securely manage AWS IAM roles, initialize SDK clients, and structure relational databases for complex property queries.

2. AI Agent Orchestration: We learned that breaking down a massive AI prompt into smaller, specialized agents (e.g., using MiniMax for conversational extraction and AWS for strict legal analysis) dramatically improves accuracy, speed, and reliability.

3. Leveraging AI to Build AI: Using Cursor completely transformed our workflow. It taught us how to effectively pair-program with AI, allowing us as 17-year-olds to build an enterprise-grade web application in a fraction of the time it would traditionally take.

Built With

• amazon-web-services
• canva
• cursor
• minimax
• supabase