Spaces:
Running
Running
| title: Worth Brain | |
| emoji: 🧠 | |
| colorFrom: blue | |
| colorTo: purple | |
| sdk: gradio | |
| python_version: "3.13" | |
| app_file: app.py | |
| pinned: false | |
| # WorthBrain | |
| Autonomous Multi-Agent Deal Intelligence System (Deal Hunter) | |
| ## Demo Video | |
| Click the link below to watch a 4-minute video for the system demo: | |
| https://drive.google.com/file/d/1HCyEESXbye_O19zVpYNPEQuyCv4hVQXy/view?usp=drive_link | |
| ## Overview | |
| WorthBrain is a multi-agent system designed to automatically discover, evaluate, and surface high value online deals (with a lot of discount). | |
| It integrates multiple AI components, including a fine-tuned open-source LLM (QLoRA), a frontier model API, and a custom neural network estimator. | |
| The system runs autonomously and streams live updates to a Gradio-based UI and mobile devices of the users by push notification. | |
| The core objective of WorthBrain is to: | |
| 1. Discover new deals online | |
| 2. Estimate fair value using multiple AI models and price data | |
| 3. Calculate discount potential | |
| 4. Select the strongest opportunity (Most valuable deal, price-wise) | |
| 5. Notify the user via an LLM-generated message and the deal table in the UI | |
| ## System Architecture | |
|  | |
| At the highest level, WorthBrain is structured into two major layers: | |
| - Agent Orchestration Layer | |
| - User Interface Layer | |
| The orchestration layer contains the Agent Framework and its sub-agents. The UI layer is responsible for rendering logs and results in real time. | |
| ### Agent Structure | |
| 1. Agent Framework | |
| - Memory | |
| - Logging | |
| - Planning Agent | |
| 2. Planning Agent coordinates: | |
| - Scanner Agent (RSS deal discovery) | |
| - Ensemble Agent (multi model price estimation, using 3 models) | |
| - Messaging Agent (LLM-generated notification) | |
| 3. The Ensemble Agent internally combines: | |
| - Specialist Agent (LLaMA 3.1 8B model Fine-tuned with QLoRA) | |
| - Frontier Agent (GPT 5 mini with RAG -- Amazon dataset) | |
| - Neural Network Agent (Custom neural network built with Scikit Learn) | |
| Each sub-agent has a clearly defined responsibility and communicates through structured objects (Deal, Opportunity, etc.) | |
| ### Runtime Flow | |
|  | |
| WorthBrain uses a producer–consumer concurrency model with queues to separate computation from UI updates. | |
| The execution flow is as follows: | |
| 1. Gradio UI triggers run_with_logging() on load. | |
| 2. run_with_logging(): | |
| - Creates log_queue and result_queue | |
| - Registers a QueueHandler for logging | |
| - Starts a background worker thread | |
| 3. Background Worker (Producer): | |
| - Executes AgentFramework.run() | |
| - During execution, logging.info(...) sends formatted log records into log_queue | |
| - After completion, places the final opportunity table into result_queue | |
| 4. stream_ui_updates() (Consumer Generator): | |
| - Continuously checks log_queue for new log messages | |
| - Appends them to persistent log state | |
| - Checks result_queue for final results | |
| - Yields updated UI state incrementally | |
| 5. Gradio renders streamed output in real time. | |
| This design prevents UI blocking while the agent system executes complex logic. | |
| ### Concurrency Model | |
| WorthBrain explicitly separates: | |
| #### Producer: | |
| - Background thread | |
| - Performs heavy computation | |
| - Writes to queues | |
| #### Consumer: | |
| - Generator loop in stream_ui_updates() | |
| - Reads from queues using get_nowait() | |
| - Streams results to UI | |
| #### Transport Layer: | |
| - queue.Queue() | |
| - Thread-safe communication | |
| This ensures proper cross-thread data exchange and continuous UI responsiveness. | |
| ## Technical Stack | |
| - Python 3.13.1 | |
| - Gradio | |
| - Plotly | |
| - LoRA fine-tuned open-source LLM | |
| - OpenAI API | |
| - Scikit Learn for custom neural network | |
| - Threading and Queue concurrency | |
| - Modal (serverless inference endpoint for the fine-tuned llm) | |
| - QLoRA (fine tuned LLaMA 3.1 8b) | |
| - logging pipeline | |
| - RSS parsing pipeline (Beautiful Soap, FeedParser etc) | |
| ## Project Scope | |
| WorthBrain is not a simple demo script. It demonstrates: | |
| - Multi-agent coordination | |
| - Model ensemble reasoning | |
| - Concurrency and background processing | |
| - Streaming UI updates | |
| - Structured system architecture | |
| The design mirrors small-scale production patterns, focusing on clarity, modularity, and separation of concerns. | |
| # WorthBrain | |
| Autonomous Multi-Agent Deal Intelligence System (Deal Hunter) | |
| ## Demo Video | |
| Click the link below to watch a 4-minute video for the system demo: | |
| https://drive.google.com/file/d/1HCyEESXbye_O19zVpYNPEQuyCv4hVQXy/view?usp=drive_link | |
| ## Overview | |
| WorthBrain is a multi-agent system designed to automatically discover, evaluate, and surface high value online deals (with a lot of discount). | |
| It integrates multiple AI components, including a fine-tuned open-source LLM (QLoRA), a frontier model API, and a custom neural network estimator. | |
| The system runs autonomously and streams live updates to a Gradio-based UI and mobile devices of the users by push notification. | |
| The core objective of WorthBrain is to: | |
| 1. Discover new deals online | |
| 2. Estimate fair value using multiple AI models and price data | |
| 3. Calculate discount potential | |
| 4. Select the strongest opportunity (Most valuable deal, price-wise) | |
| 5. Notify the user via an LLM-generated message and the deal table in the UI | |
| ## System Architecture | |
|  | |
| At the highest level, WorthBrain is structured into two major layers: | |
| - Agent Orchestration Layer | |
| - User Interface Layer | |
| The orchestration layer contains the Agent Framework and its sub-agents. The UI layer is responsible for rendering logs and results in real time. | |
| ### Agent Structure | |
| 1. Agent Framework | |
| - Memory | |
| - Logging | |
| - Planning Agent | |
| 2. Planning Agent coordinates: | |
| - Scanner Agent (RSS deal discovery) | |
| - Ensemble Agent (multi model price estimation, using 3 models) | |
| - Messaging Agent (LLM-generated notification) | |
| 3. The Ensemble Agent internally combines: | |
| - Specialist Agent (LLaMA 3.1 8B model Fine-tuned with QLoRA) | |
| - Frontier Agent (GPT 5 mini with RAG -- Amazon dataset) | |
| - Neural Network Agent (Custom neural network built with Scikit Learn) | |
| Each sub-agent has a clearly defined responsibility and communicates through structured objects (Deal, Opportunity, etc.) | |
| ### Runtime Flow | |
|  | |
| WorthBrain uses a producer–consumer concurrency model with queues to separate computation from UI updates. | |
| The execution flow is as follows: | |
| 1. Gradio UI triggers run_with_logging() on load. | |
| 2. run_with_logging(): | |
| - Creates log_queue and result_queue | |
| - Registers a QueueHandler for logging | |
| - Starts a background worker thread | |
| 3. Background Worker (Producer): | |
| - Executes AgentFramework.run() | |
| - During execution, logging.info(...) sends formatted log records into log_queue | |
| - After completion, places the final opportunity table into result_queue | |
| 4. stream_ui_updates() (Consumer Generator): | |
| - Continuously checks log_queue for new log messages | |
| - Appends them to persistent log state | |
| - Checks result_queue for final results | |
| - Yields updated UI state incrementally | |
| 5. Gradio renders streamed output in real time. | |
| This design prevents UI blocking while the agent system executes complex logic. | |
| ### Concurrency Model | |
| WorthBrain explicitly separates: | |
| #### Producer: | |
| - Background thread | |
| - Performs heavy computation | |
| - Writes to queues | |
| #### Consumer: | |
| - Generator loop in stream_ui_updates() | |
| - Reads from queues using get_nowait() | |
| - Streams results to UI | |
| #### Transport Layer: | |
| - queue.Queue() | |
| - Thread-safe communication | |
| This ensures proper cross-thread data exchange and continuous UI responsiveness. | |
| ## Technical Stack | |
| - Python 3.13.1 | |
| - Gradio | |
| - Plotly | |
| - LoRA fine-tuned open-source LLM | |
| - OpenAI API | |
| - Scikit Learn for custom neural network | |
| - Threading and Queue concurrency | |
| - Modal (serverless inference endpoint for the fine-tuned llm) | |
| - QLoRA (fine tuned LLaMA 3.1 8b) | |
| - logging pipeline | |
| - RSS parsing pipeline (Beautiful Soap, FeedParser etc) | |
| ## Project Scope | |
| WorthBrain is not a simple demo script. It demonstrates: | |
| - Multi-agent coordination | |
| - Model ensemble reasoning | |
| - Concurrency and background processing | |
| - Streaming UI updates | |
| - Structured system architecture | |
| The design mirrors small-scale production patterns, focusing on clarity, modularity, and separation of concerns. |