My Journey in AI/ML: From Embedded Systems to RAG-based LLMs
Introduction
My journey from embedded systems engineering to AI/ML development has been fascinating and unexpected. What started as a passion for automotive embedded systems has evolved into developing cutting-edge AI solutions, particularly RAG-based Large Language Models. Here’s how my technical evolution unfolded and what I’ve learned along the way.
The Foundation: Embedded Systems
My technical journey began with a deep dive into embedded systems during my Master’s degree in Automotive Embedded Systems from ESIGELEC (France) and MAHE (India). This foundation taught me:
Core Principles
- Real-time Constraints: Understanding the critical importance of timing in systems
- Resource Optimization: Working with limited memory and processing power
- Safety-Critical Thinking: Developing software where failures have real-world consequences
- Hardware-Software Integration: Bridging the gap between physical and digital worlds
Technical Skills Acquired
Working with automotive data analysis, implementing machine learning pipelines, and ensuring data quality gave me a solid foundation in systematic, robust software development. I gained experience in Python-based data processing, sensor data analysis, and anomaly detection using libraries like NumPy, Pandas, and Scikit-learn.
The Transition: Computer Vision and Robotics
During my early career, I discovered the exciting world of computer vision and robotics. This was my first real exposure to AI/ML technologies:
Key Projects
- Computer Vision with TensorFlow: Implemented CNN-based image classification for robot navigation
- Path Planning: Developed algorithms for real-time robot navigation using reinforcement learning
- Speed Control: Developed Kalman filter-based speed control systems for agriculture robots
- Sensor Fusion: Combined camera for robust perception using deep learning
Technical Growth
I implemented CNN-based image classification for robot navigation, developed real-time path planning algorithms, and worked on sensor fusion combining camera data. This experience taught me the power of deep learning in solving real-world problems and sparked my interest in pursuing AI/ML further.
Key Achievements:
- Developed obstacle detection and classification systems using TensorFlow
- Implemented real-time image processing pipelines with OpenCV
- Built robust perception systems for autonomous navigation
- Applied convolutional neural networks for visual recognition tasks
The Evolution: Python Development
Transitioning from embedded C to Python development marked a significant shift in my career, focusing on:
Technical Responsibilities
- Contributing to the Python ecosystem for simulation software, its maintenance, and enhancement
- Creating scripts and tools to automate simulation workflows
- Building user-friendly themes and plugins for documentation automation
- Implementing automated testing and deployment pipelines
- Building user-friendly interfaces for complex simulation tools
Learning Python Ecosystem
The transition to Python development opened up new possibilities. I learned modern development practices including asynchronous programming, comprehensive testing with pytest, type hinting for better code quality, and API development patterns. This experience taught me how to build scalable, maintainable software for engineering applications.
The Current Focus: RAG-based LLMs
My latest adventure involves developing Retrieval-Augmented Generation (RAG) based Large Language Models. This represents the convergence of all my previous experiences:
Why RAG?
RAG combines the power of large language models with domain-specific knowledge retrieval, making it perfect for technical documentation and knowledge management - areas where precision and accuracy are crucial.
Technical Implementation
I implemented a comprehensive RAG system using LangChain, OpenAI embeddings, and vector stores. The system processes technical documentation, creates semantic embeddings, and provides accurate question-answering capabilities for engineering teams.
Key Components:
- Document processing and chunking for optimal retrieval
- Vector embeddings using OpenAI’s embedding models
- FAISS vector store for efficient similarity search
- Retrieval-augmented generation chain for accurate responses
- Source document tracking for transparency and verification
Key Challenges and Solutions
1. Data Quality and Preprocessing
- Challenge: Technical documentation varies in format and quality
- Solution: Implemented robust preprocessing pipelines with document parsing and cleaning
2. Embedding Quality
- Challenge: Generic embeddings don’t capture domain-specific semantics
- Solution: Fine-tuned domain-specific embedding models
3. Retrieval Accuracy
- Challenge: Ensuring retrieved documents are truly relevant
- Solution: Implemented hybrid search combining semantic and keyword-based retrieval
4. Response Quality
- Challenge: Generating accurate, helpful responses
- Solution: Prompt engineering and response validation mechanisms
Lessons Learned
Technical Insights
- Foundation Matters: My embedded systems background taught me to think about constraints and optimization - crucial for efficient AI systems
- Iterative Development: The transition from waterfall (embedded) to agile (AI/ML) development methodologies
- Community and Open Source: The Python/AI community’s collaborative approach vs. proprietary embedded development
Career Development
- Continuous Learning: The AI field evolves rapidly; staying current requires constant learning
- Interdisciplinary Skills: Combining domain expertise (automotive/engineering) with AI capabilities creates unique value
- Problem-First Approach: Focus on solving real problems rather than applying trendy technologies
Future Directions
Looking ahead, I’m excited about several emerging areas:
Edge AI
Bringing AI capabilities to embedded systems - combining my original passion with current expertise. I’m exploring model optimization techniques like quantization and pruning to deploy TensorFlow models on resource-constrained devices, enabling real-time inference in automotive and IoT applications.
Automotive AI
The intersection of my automotive background and AI expertise opens exciting possibilities:
Current Focus Areas: - Autonomous driving algorithms using computer vision and deep learning - Predictive maintenance systems with time series analysis and sensor data - Advanced driver assistance systems (ADAS) with real-time inference capabilities - Lane detection and obstacle recognition using convolutional neural networks - Vehicle health monitoring through IoT sensors and machine learning models
Key Achievements: - Developed lane detection systems using CNN architectures - Built predictive maintenance models using LSTM networks for time series analysis - Implemented real-time inference systems for automotive safety applications
Advanced RAG Systems
Advanced Capabilities I’m Exploring:
- Multimodal RAG systems combining text, images, and diagrams
- Real-time knowledge updating and incremental learning
- Domain-specific fine-tuning for specialized engineering contexts
- Conversational AI with memory for complex technical discussions
- Local deployment using HuggingFace models for privacy-sensitive applications
- Guard rails and ethical considerations in AI responses
Conclusion
The journey from embedded systems to AI/ML has been challenging but incredibly rewarding. Each phase has built upon the previous one:
- Embedded Systems: Taught me precision, constraints, and systematic thinking
- Computer Vision: Introduced me to the power of machine learning
- Python Development: Provided tools and ecosystem knowledge
- RAG-based LLMs: Combines everything into solving complex, real-world problems
Key Insights
The key insight is that diverse technical backgrounds are actually advantageous in AI/ML. The discipline from embedded systems, the practical focus from robotics, and the software engineering practices from Python development all contribute to building better AI systems.
Looking Forward
The future lies at the intersection of traditional engineering disciplines and AI. As someone who’s navigated this path, I’m excited to see where this journey leads next!