Inspiration We noticed that millions of Micro, Small, and Medium Enterprises (MSMEs) are the backbone of the economy but remain financially "invisible" due to a lack of formal statements. Simultaneously, banks face a "Last Mile" problem in Green Finance: the cost to verify the environmental impact of small loans is too high. We asked: What if AI could turn a vendor's "Dark Data"—like utility bills—into verifiable "Carbon Assets"? This drove us to build GreenCredit Agent.

What it does GreenCredit Agent is an autonomous AI system serving two roles:

For Vendors: A 24/7 Sustainability CFO that analyzes bills via OCR, calculates carbon footprints, and offers voice-based advice on saving energy costs. For Banks: An Automated Green Risk Officer that converts unstructured data into a standardized Green Credit Score, enabling lower-risk, high-impact lending. How we built it We built an Agentic Architecture with three layers:

Tech Stack: Used LangChain for agent autonomy, GPT-4o/DeepSeek for reasoning, and Pinecone (Vector DB) for RAG-based knowledge retrieval. Core Algorithm: We developed a Carbon-Credit Fusion Model to quantify trustworthiness. The Green Credit Score ($S_{green}$) is calculated as: S g r e e

n

α ⋅ F ( C a s h F l o w ) + β ⋅ G ( E n e r g y E f f i c i e n c y ) + γ ⋅ C ( C o n s i s t e n c y ) S green ​ =α⋅F(CashFlow)+β⋅G(EnergyEfficiency)+γ⋅C(Consistency) Where $\mathcal{G}$ represents the Green Index derived from the ratio of revenue to carbon emissions. Challenges we ran into Dirty Data: Real-world receipts are messy. We had to fine-tune prompt engineering to help the LLM infer context from partial OCR results. Hallucinations: To prevent the agent from inventing financial subsidies, we implemented a strict RAG pipeline, forcing it to cite specific policy documents. Privacy: We built a Data Desensitization Layer to mask PII before data enters the LLM context window. Accomplishments that we're proud of Voice-First Interface: We ensured inclusivity by allowing users to interact via voice notes, lowering the barrier for non-tech-savvy vendors. Proactive Nudging: The agent doesn't just wait for queries; it actively alerts users about anomalies (e.g., energy spikes) to save costs. What we learned Agentic AI: The true power of LLMs lies in their ability to act, connecting perception (OCR) to action (API calls), not just chatting. Empathy in Design: Building for MSMEs taught us that the best technology is invisible; complex UIs are a barrier to adoption. What's next Phase 2 (Jan-Feb 2025): Refine the scoring algorithm with synthetic data. Phase 3 (March 2025): Live demo of the full "Upload-to-Loan" workflow at Innovation Day. Long-term: Integration with IoT smart meters for fully automated data collection.

Built With

• deepseek
• langchain
• openai-api
• python