Repositories: doc-agent (Backend) & doc-agent-fe (Frontend)

📌 Project Overview

Engineered a production-ready Document Retrieval-Augmented Generation (RAG) system designed to ingest, process, and interrogate complex document structures. The platform enables users to upload dense materials and perform highly accurate, context-aware semantic searches through an interactive conversational interface.

🛠️ The Challenge

Enterprise documentation rarely exists in clean, plain text. Extracting actionable insights from multi-page PDFs, nested tables, and dense contractual clauses is a major bottleneck. Standard keyword search fails to understand intent, and naive LLM approaches often hallucinate or lose context when dealing with large, complex, unstructured documents.

💡 Technical Solution

1. Robust Ingestion & Indexing Engine (doc-agent)

• Complex PDF Processing: The backend is specifically optimized to parse and chunk difficult document structures, ensuring that tabular pricing data and dense text (such as legal Terms & Conditions) are broken down semantically without losing their structural integrity.

• Vector Search Infrastructure: Document embeddings are generated and indexed directly into a Qdrant vector database, serving as the high-performance retrieval engine for complex semantic queries.

2. Interactive Web Application (doc-agent-fe)

• Responsive UI: A dedicated front-end companion application built to provide a frictionless user experience for document management.

• Unified Upload & Chat Interface: Users can easily upload their complex PDFs directly into the system and immediately begin querying the ingested material via a real-time chat interface.

📊 Key Features & Capabilities

• Context-Aware Interrogation: The RAG pipeline ensures that every answer generated is strictly grounded in the ingested documents, providing accurate, traceable responses to user queries.

• Tabular & Legal Data Handling: Specifically tuned to handle the nuances of pricing tables and legal fine print, mitigating the common pitfalls of standard text chunking.

• Decoupled Architecture: A clean separation of concerns, isolating the heavy computational lifting of document processing and vector storage from real-time client interactions.

⚙️ Core Architecture Focus

• Backend & Data Pipeline: doc-agent (Document Parsing, Vectorization, RAG Orchestration)

• Vector Database: Qdrant

• Frontend Client: doc-agent-fe (Responsive Web App, Chat UI, File Management)

  Front End code

Hierarchical Dual-Processor Hardware & ROS 2 Autonomy Architecture

📌 Project Overview

Designed and built a modular, differential-drive Autonomous Mobile Robot (AMR) chassis from the ground up to solve critical computational bottlenecks in dynamic indoor navigation. By decoupling non-deterministic high-level spatial perception (SLAM, path planning) from hard real-time low-level motor control, the system eliminates control instabilities caused by operating system latency.

🛠️ The Challenge

Modern AMRs must process massive point-cloud and vision datasets while simultaneously running high-frequency motor control loops. Executing both on a single processor or under a non-deterministic OS (like standard Linux) leads to thread contention, latency spikes, and severe control loop instability.

💡 Technical Solution

1. Dual-Processor Hierarchical Architecture

• High-Level Compute: Raspberry Pi 5 handling spatial mapping, trajectory planning, and high-level behavior trees.

• Low-Level Deterministic Control: STM32G431RB microcontroller executing real-time timer-interrupt control loops and hardware quadrature decoding.

• Inter-Process Transport: High-speed, asynchronous UART communication bridge integrated with ROS 2 DDS transport nodes.

2. Custom Power & Motor Controller PCB

• High-Efficiency Power Distribution: Custom PCB featuring MP2338 buck converters delivering >90% energy efficiency, effectively isolating heavy computational spikes from motor actuation to prevent voltage sags.

• Integrated Motor Drivers: Embedded DRV8231 ICs providing hardware-level overcurrent protection and direct PWM velocity control.

3. Real-Time Embedded Control & Filtering

• Discrete PID Loop: Fixed-frequency velocity control algorithm executing on STM32 hardware timers, achieving a rapid 100ms step-response time with minimal overshoot under dynamic physical loads.

• Signal Conditioning: Implemented an Exponential Moving Average (EMA) filter on 4-channel quadrature encoder feedback to attenuate high-frequency quantization noise without introducing lag.

4. Full ROS 2 Autonomy Stack

• Graph-Based SLAM: Integrated slam_toolbox with an RPLIDAR C1 sensor for real-time 2D occupancy grid generation.

• Navigation & Path Planning: Implemented the ROS 2 Nav2 framework featuring Behavior Trees (BT Navigator) and dynamic multi-layered costmaps (Static, Obstacle, Inflation).

📊 Key Performance Metrics & Results

• Mapping Accuracy: Generated 5 cm/pixel resolution maps with <2 cm perimeter deviation across a $10\text{m} \times 15\text{m}$ indoor test environment.

• Deterministic Stability: Achieved 100% decoupling of high-frequency motor loops from high-level CPU load spikes.

• Power Efficiency: Maintained >90% power conversion efficiency across peak operational loads.

⚙️ Tech Stack & Hardware Specs

• Processors: STM32G431RB (ARM Cortex-M4 with hardware FPU), Raspberry Pi 5

• Software Frameworks: ROS 2 Jazzy, Nav2, slam_toolbox, C/C++, Python

• Electronics & Power: Custom Buck Converters (MP2338), H-Bridge Motor Drivers (DRV8231)

• Kinematics & Sensors: Differential Drive Configuration, RPLIDAR C1, 4-Channel Quadrature Encoders

  Robot C/C++ code for naivigation and mapping
  Motor controller code for STM32

  PCB Schematic