Week 15 – Exploration and Advanced Learning in Generative AI
Dates: September 7 – September 13
Internship: AI/ML Intern at SynerSense Pvt. Ltd.
Mentor: Praveen Kulkarni Sir
Focus
Following the completion of the first fine-tuning and benchmarking phase, this week was dedicated to exploration and independent learning. With no immediate assignments from the mentor, I utilized this time to deepen my understanding of Deep Learning architectures, Retrieval-Augmented Generation (RAG) systems, and the evolving landscape of Generative AI tools and frameworks.
This self-directed exploration aimed to strengthen theoretical knowledge and identify potential improvements for future AI-driven solutions within the project domain.
Goals for the Week
- Study the internal mechanisms of Deep Learning models, focusing on transformer architectures and fine-tuning strategies
- Explore Retrieval-Augmented Generation (RAG) and its integration with LLMs for factual, context-aware responses
- Experiment with emerging Generative AI tools and APIs
- Review OpenAI and Hugging Face documentation for model deployment, dataset structuring, and evaluation techniques
- Summarize learnings into reusable notes for future reference
Tasks Completed
| Task | Status | Notes |
|---|---|---|
| Explored transformer architectures and attention mechanisms | ✅ Completed | Reviewed “Attention Is All You Need” and transformer visualizations |
| Studied RAG (Retrieval-Augmented Generation) pipelines | ✅ Completed | Implemented a small RAG demo combining FAISS retriever and OpenAI embeddings |
| Tested Generative AI APIs (OpenAI, Hugging Face, Replicate) | ✅ Completed | Compared text and image generation outputs across platforms |
| Researched evaluation metrics for generative models | ✅ Completed | Focused on BLEU, ROUGE, and human evaluation consistency |
| Documented notes and shared insights | ✅ Completed | Summarized findings into markdown-based technical notes for the team |
Key Learnings
-
Transformers form the backbone of modern AI.
Understanding the role of self-attention and token embeddings clarified how large language models generalize patterns and contextual cues. -
RAG enhances reliability in LLMs.
Retrieval-Augmented Generation bridges the gap between generative capabilities and factual grounding by injecting real-time contextual data into prompts. -
Evaluation remains an open challenge.
Unlike classification models, generative outputs require both quantitative (BLEU, ROUGE) and qualitative (human judgment) evaluation to assess coherence and factuality. -
The ecosystem is evolving rapidly.
Tools like LangChain, LlamaIndex, and OpenAI’s updated APIs simplify building applied AI solutions, but they also demand continuous learning.
Challenges and Solutions
| Challenge | Resolution |
|---|---|
| Difficulty in implementing RAG without proper vector storage | Integrated FAISS as a lightweight and efficient retriever |
| Understanding transformer attention visualization | Used open-source visualizers and Hugging Face course examples |
| Managing resource limits for model testing | Shifted computation to Google Colab Pro and reduced sample size |
References
- Attention Is All You Need (Vaswani et al., 2017)
- Retrieval-Augmented Generation (RAG) – Hugging Face Docs
- LangChain Documentation
- FAISS – Facebook AI Similarity Search
- OpenAI API Documentation
Goals for Next Week
- Consolidate learnings into a mini project or internal demo showcasing RAG or transformer visualization
- Document best practices for integrating retrieval and fine-tuning in production workflows
- Begin preparing the final internship report and presentation slides
Screenshots (Optional)
Screenshot of the RAG demo showing the retriever fetching top-ranked context passages and the model generating context-aware responses.
“Week 15 served as a bridge between execution and understanding — a period of reflection and learning that strengthened the theoretical and practical grasp of next-generation AI systems.”