FUNDS.AI

Inspiration

We noticed a critical inefficiency in the non-profit sector: while the OurSG Grants portal offers vital funding, finding the right grant is a manual, exhausting process. With frequent updates and a rigid search interface utilising up to 40 different filters, non-profits were suffering from "search fatigue."

We realised organisations were spending countless hours sifting through irrelevant information—time that should be spent on their mission. We wanted to build a solution that didn't just list grants, but actively "pulled" the right information to the users who needed it, ensuring no funding opportunity is missed due to a lack of manpower.

What it does

FUNDS.AI is an intelligent enhancement layer for the OurSG Grants portal that fully automates the grant discovery process.

• ⁠Automated Scraping & Monitoring: Our system runs a daily Cronjob to scrape the official portal, ensuring our repository is always synchronised with the latest government data. ⁠ ⁠*Semantic Search: We replaced rigid checkboxes with a Vector Database enabling natural language search. Users can simply describe their organization (e.g., "I'm a tech startup helping the elderly"), and our AI understands the context to find relevant matches.
• ⁠AI Grant Advisor: Every grant page features an embedded RAG-powered chatbot. It acts as a dedicated consultant, answering specific questions about eligibility, KPIs, and funding amounts in plain English.
• ⁠Proactive Alerts: Users don't need to log in daily. Our system sends real-time email notifications whenever a new grant matches their specific profile and tags.

How we built it

We combined modern web technologies with an advanced AI retrieval pipeline:

-⁠ ⁠Data Pipeline: We built a custom Python scraper (using Selenium) to extract and normalize unstructured data from the OurSG portal. -⁠ ⁠Backend & AI: We utilised a Vector Database (via Prisma and PostgreSQL) to store grant information as embeddings. This allows our RAG (Retrieval Augmented Generation) system to perform semantic searches based on intent rather than just keywords.

• ⁠Frontend: Built with Next.js, the application provides a clean, responsive dashboard where users can manage their notifications and interact with the AI agents.
• ⁠AI Integration: We leveraged Google Gemini to power our "Agentic Assistant," which classifies unstructured grant descriptions into standardised categories and drives the chatbot's reasoning capabilities.

Challenges we ran into

-⁠ ⁠Taming Unstructured Data: Government grant portals often use complex HTML structures and inconsistent formatting. Converting this "messy" raw data into structured, clean JSON that our database and AI could reliably process was a significant hurdle. -⁠ ⁠Preventing AI Hallucinations: When building the "Grant Advisor" chatbot, we faced the risk of the AI inventing eligibility criteria. We had to rigorously tune our system prompts and implement strict Retrieval Augmented Generation (RAG) protocols to ensure the AI only answered based on the retrieved facts, not its own creative tendencies. -⁠ ⁠Latency vs. Accuracy: Performing vector similarity searches and generating AI responses in real-time can be slow. We had to optimise our embedding generation pipeline to ensure the user experience remained snappy without sacrificing the depth of the AI's analysis.

Accomplishments that we're proud of

-⁠ ⁠Scalable Architecture: We built a system that requires no initial monetary investment and can scale to handle thousands of grants with negligible cost increases, thanks to our optimised vector search implementation.

• ⁠Democratised Guidance: By embedding intelligent chatbots on every grant page, we effectively gave every non-profit a free "consultant," providing instant clarification on complex criteria that usually requires a phone call or email to clarify.
• ⁠End-to-End Automation: Successfully linking the scraper, the vector database, and the notification engine means the platform runs itself—grants are found, processed, and delivered to users without human intervention.

What we learned

We gained a deep appreciation for Retrieval Augmented Generation (RAG). We learned that for specialised knowledge (like government grants), a generic LLM isn't enough; it needs a reliable "knowledge base" to prevent hallucinations. We also learned the distinct strengths of vector databases versus relational databases when handling the semantic nuance of real-world queries.

What's next for FUNDS.AI

-⁠ AI-Drafted Proposals: Moving beyond discovery, we plan to implement an "Auto-Draft" agent. Users will be able to upload their company profile, and the AI will generate a first draft of the grant application based on the specific requirements of the selected grant.

• Predictive Success Scoring: We aim to analyse historical data to provide users with a "Success Probability" score, helping them prioritise grants they are most likely to win. -⁠ Multi-Source Aggregation: While we currently focus on OurSG Grants, the architecture is ready to expand. We plan to aggregate private foundation grants and international funding, becoming the single source of truth for all non-profit funding.

Built With

• betterauth
• neon
• nextjs
• pgvector
• postgresql
• prisma
• pydantic
• python
• selenium
• typescript