Project Overview

The coffee industry is rich with complexity, creating a significant information gap for both consumers and aspiring baristas. PIKOPI was conceived to bridge this gap by creating a centralized, AI-powered platform. Our team of four developed a comprehensive solution that addresses key user needs: identifying coffee quality, predicting its potential, finding the perfect flavor profile, and getting expert advice on demand. The result is a robust web application that serves as an educator, a recommender, and a technical assistant.

My Role & Contribution: The Agentic Chatbot and Unified UI

While this was a collaborative effort, my focus was on two critical components: building the conversational AI brain and crafting the user-facing application that brought all our work together.

1. The Agentic AI Chatbot (The "Brain")

I engineered an advanced conversational agent, moving beyond a simple Q&A bot to create a true Agentic AI capable of autonomous reasoning and tool use.

• Architecture: I used LangGraph to build a cyclic workflow, allowing the agent to think, act, and reflect. The core LLM was Google Gemini 2.5 Flash.
• Intelligent Tools: I developed and integrated three distinct tools for the agent to use:
  • RAG Knowledge Base: A retrieve_coffee_knowledge tool connected to a Qdrant Vector Database, enabling the agent to answer theoretical questions about coffee accurately and without hallucination.
  • Real-Time Location Finder: A find_cafes_with_maps tool that uses Google Maps Grounding. It's context-aware (knows the current time to find open cafes) and robust against user typos thanks to vector search.
  • Smart Brew Calculator: A calculate_brew_recipe tool with deterministic math logic to provide precise brewing ratios for various methods (V60, Espresso, etc.), a task where LLMs often fail.
• Prompt Engineering & Persona: I designed a comprehensive System Prompt to give the agent the personality of a friendly, expert barista. This involved using Few-Shot Prompting (providing example dialogues) and strict Guardrails to ensure it stayed on topic and preserved critical data like Google Maps links in its responses.

2. The Unified Streamlit Application (The "Face")

I was solely responsible for developing the entire frontend and integrating all backend modules into a seamless user experience.

• Framework: I chose Streamlit for its rapid development capabilities and powerful data app features.
• UI Development: I built the complete user interface, including the modern chat interface, a functional sidebar, and custom CSS for a polished, branded look.
• System Integration: My most crucial role was integrating the work of all four team members. I connected Kadek's classification model to a file uploader, Alivia's prediction model to an input form, and Riko's personalization logic into the recommendation flow. My chatbot served as the central conversational hub for the entire platform.
• User Experience: I implemented features like Session State to maintain chat history and a typing effect for AI responses to create a more dynamic and engaging interaction.

Team Collaboration & Integrated Features

Our platform's rich functionality was made possible by the collective effort of the entire team. I had the pleasure of integrating the excellent work of my colleagues:

• Coffee Bean Classification (by I Kadek Rangga Sandi K.W): Kadek developed a Deep Learning model using EfficientNetB3 to classify coffee bean images into four categories (Defect, Peaberry, etc.), providing users with instant visual quality assessment.
• Cupping Score Prediction (by Alivia Azizah): Alivia built a Machine Learning model to predict a coffee's potential cupping score based on its physical and sensory attributes, offering a quantitative measure of quality.
• Personalized Recommendations (by Riko Abiyasa Dwi Andika): Riko implemented a RAG-based system that matches a user's flavor preferences (e.g., "fruity," "chocolatey") with a database of coffee profiles to deliver personalized bean recommendations.

Tech Stack

• Frontend: Streamlit
• AI Orchestration & Backend: LangChain, LangGraph, Python
• LLM & AI Services: Google Gemini 2.5 Flash, Google Maps Platform, Google Generative AI Embeddings
• Database: Qdrant Cloud (Vector Database)
• Core Libraries: Pandas, Pytest, Pytz, TensorFlow/Keras

Outcome & Reflection

The project culminated in a fully functional, deployed web application that successfully demonstrates the power of combining different AI disciplines. My work on the agent and UI taught me invaluable lessons in system architecture, API integration, and the art of creating a user-centric AI product. It was a fantastic experience in technical execution and, most importantly, in team collaboration.