These projects are not about replicating state-of-the-art benchmarks or shipping polished end products.

The goal of every repository here is to build from scratch after genuinely understanding the underlying concepts — and then to be honest about where things break, what the model cannot do, and why. Limitations are documented as carefully as results, because understanding failure modes is how real learning happens.

This matters especially coming from a Mechanical Engineering background: the instinct here is not to chase accuracy numbers, but to ask why does this architecture work, where does it fail, and what does that tell us about the problem? That question drives every project in this portfolio.

R-B.A.T (RAG-Based Academic Tutor & Assessment Console) is my long-term effort to build an institution-ready academic AI system that goes beyond chatbot-style interactions.

The project explores how Retrieval-Augmented Generation, local LLMs, structured workflows, and educational guardrails can be combined to support the complete teaching-learning cycle—from content delivery and student tutoring to assessment creation and academic resource generation.

Designed around real institutional constraints such as data privacy, limited hardware, and zero cloud dependency, the entire system runs locally using open-source models through Ollama.

🎓 Tutor Mode Grounded question answering over prescribed textbooks with syllabus-aware retrieval, diagram support, and multiple explanation styles.

📝 Assessment Mode Automatic generation of Internal and Semester Examination question papers with Bloom's Taxonomy alignment, CO mapping, and institutional formatting.

📖 Evaluation Mode Generation of textbook-grounded model answers and marking schemes for faculty reference and student preparation.

📊 Presentation Mode Automated lecture presentation generation using RAG-derived content, diagrams, flowcharts, and customizable academic themes.

Current focus: evolving R-B.A.T from a textbook-grounded tutor into a comprehensive academic co-pilot for teaching, assessment, and classroom content generation.

Classical Machine Learning  →  Neural Networks & Deep Learning
        ↓                               ↓
Representation Learning     →  Generative & Multimodal AI
                                       ↓
              Agentic AI Systems & LLM Applications
                            ↓
              AI for Engineering Applications

Each project is packaged with a report, notebook, README, requirements, and reproducible workflow.

Classical ML · Deep Learning · Representation Learning — the ground every later project is built on.

A concept-driven demonstration of why linear models fail and why hidden layers are necessary. Walks through OR and AND (linearly separable), breaks logistic regression on XOR to show where it fails — then solves it with a single hidden layer MLP. The focus is entirely on decision boundaries and architectural necessity, not accuracy.

Logistic Regression MLP Decision Boundaries

Tokenization · Embeddings · Transformers · LoRA · QLoRA

End-to-end Word2Vec (Skip-Gram + Negative Sampling) built from scratch in PyTorch, extended into an interactive browser-based embedding explorer. Exports intermediate checkpoints across epochs and visualises how semantic structure gradually emerges from random vectors — nearest neighbours, similarity scoring, analogy solving, geometric clustering.

Built a custom Byte Pair Encoding (BPE) tokenizer from scratch using the WikiText-2 corpus and extended it into an interactive browser-based tokenizer visualizer. The project explores corpus cleaning, vocabulary learning, subword formation, rare-word decomposition, compression efficiency, and tokenizer evaluation while demonstrating how modern NLP systems represent language through reusable subword units.

A hands-on exploration of Parameter-Efficient Fine-Tuning (PEFT), progressing from LoRA-based adaptation of BERT to QLoRA-based instruction tuning of Gemma 2B. The project investigates how large language models can be adapted by training only a tiny fraction of their parameters while significantly reducing memory requirements through 4-bit quantization.

What makes this different: the repository documents the complete engineering journey — including an attempted QLoRA implementation on BERT, debugging of bitsandbytes compatibility issues, architectural analysis of encoder vs decoder models, and a successful migration to Gemma 2B. Rather than hiding failed experiments, the project preserves them as learning artifacts.

Multi-agent orchestration · LangGraph · CrewAI Flows · Reflection loops · Persistent memory

A memory-driven multi-agent research system that transforms a user query into an evidence-backed research report through planning, retrieval, critique, reflection, and persistent learning. Built with LangGraph, D.A.R.I.A. combines domain-specific RAG, web research, evidence curation, a dedicated Research Critic, and structured summarization to move beyond traditional one-shot question answering.

What makes this different: retrieved evidence is independently evaluated by a Research Critic that identifies missing concepts, generates improvement feedback, and triggers replanning when research quality is insufficient. Research plans, information gaps, critic feedback, and final reports are stored as persistent memory and reused in future investigations.

A CrewAI Flow-powered logistics decision intelligence platform that transforms supply chain metrics into executive-level optimization playbooks. Parallel analytical branches (inventory + logistics) synchronize through Flow barriers, pass into a memory-aware strategist, and every strategy is reviewed by an independent Critic Agent before a revision cycle triggers when necessary.

What makes this different: historical optimization playbooks are actively retrieved and injected into future strategy generation — delta reasoning where recommendations evolve from previous decisions rather than restarting from scratch.

A multi-agent financial intelligence platform that transforms bill images into actionable financial insights through deterministic analytics, expense forecasting, and collaborative AI reasoning. Built with AutoGen AgentChat, the system combines OCR, spending analytics, recurring expense detection, and a dedicated Advisor–Reviewer collaboration workflow to produce explainable, verified financial recommendations.

What makes this different: deterministic financial analytics are deliberately separated from LLM reasoning. Structured financial reports are generated first, then independently reviewed by a Financial Reviewer before the Financial Advisor produces the final verified consultation. Every recommendation is supported by structured analytics and accompanied by an explainable consultation trace.

A LangGraph ReAct competitor intelligence platform for clothing stores — built in 2 days. Discovers nearby competitors via Apify, enriches with BestTime traffic data (with an inference fallback for missing coverage), runs a reflection-driven validation loop, and stores every completed analysis in ChromaDB as a queryable RAG assistant.

What makes this different: the dual-layer traffic strategy (empirical API → inference fallback) guarantees 100% traffic coverage even when primary data is unavailable.

A confidence-scored reflection loop for travel planning: after initial data collection, the agent evaluates its own information completeness (0–100%), identifies knowledge gaps, and re-searches with targeted queries before generating the final guide. Maximum 2 re-search cycles to prevent runaway API consumption. Every cycle's verdict and gaps are surfaced in an Agent Insights tab — internal reasoning fully auditable.

Autonomous research agent built on the ReAct paradigm from scratch — no LangChain, no LangGraph, no framework. Full Thought → Action → Observation → Summary loop per research question using a multi-LLM split: Groq LLaMA 3.3 70B for reasoning steps, Gemini Flash Lite for planning and synthesis.

Honest implementation: limitations (single-pass loop, in-session memory only, no reflection) are documented as the natural next improvements, not hidden.

Gemini Groq LLaMA 3.3 70B Tavily ReAct Pattern Streamlit

Four specialised agents — Research (RAG) → Image (diffusion) → Reviewer → Manager — orchestrated to transform a topic prompt into structured educational content. Failure cases (handoff breakdowns, irrelevant RAG retrievals, divergent diffusion diagrams) documented alongside the working pipeline.

CrewAI Gemini Groq FLUX RAG Streamlit

First agentic implementation. Built as a live demo for the final session of a 3-day AI workshop (SNS College of Technology, Jan 2026). Replaces LLM routing with a FAISS L2 distance threshold — one fewer API call per query, fully deterministic, and makes the agent's decision logic transparent to a non-CS audience.

Gemini API SentenceTransformers FAISS pypdf Streamlit

RAG pipelines · Multimodal systems · Generative models · Vision-Language

A multimodal RAG pipeline handling text, tables, and figures from academic PDFs in a single unified system. Docling preserves structure across all three modalities; BGE embeds into FAISS; Gemini answers with vision-enabled summarisation for figures and tables. Queries each modality separately and merges before generation.

×4 image reconstruction (SRGAN and RRDB-based ESRGAN) built from scratch with patch-based training, warm-up stability phases, and adversarial fine-tuning on DIV2K. Results under constrained compute are documented honestly — mode collapse, discriminator instability, and the gap between perceptual loss and PSNR are all explained rather than cherry-picked around.

SRGAN ESRGAN RRDB VGG Perceptual Loss DIV2K

Caption generation progressing from CNN–LSTM baseline (InceptionV3 + LSTM) through Transformer decoder to pretrained vision–language transformers. Validates learning via controlled overfitting before scaling up. Documents why pretrained models outperform naive fine-tuning on limited data — the analysis of what breaks is as central as the results.

InceptionV3 LSTM Transformer Beam Search HuggingFace

• Building LLM foundations and applied NLP systems, from tokenization and embeddings to parameter-efficient fine-tuning
• Developing RAG, GraphRAG, and agentic AI applications for education, research, and decision support
• Applying AI to engineering, manufacturing, logistics, and supply chain optimization
• Creating interactive AI learning experiences, workshops, and educational tools for students and faculty
• Exploring local and hybrid LLM deployments (Gemma, Ollama, Groq, Gemini) for scalable and cost-efficient AI systems
• Designing multi-agent and multi-LLM architectures that combine specialized models, tools, and retrieval workflows

R. Ruthuraraj · Assistant Professor · Mechanical Engineering · SNS College of Technology AICTE QIP Programme — AI to Generative AI, IIIT Allahabad

This portfolio documents a self-directed learning journey from classical machine learning and statistical modelling to deep learning, generative AI, retrieval-augmented generation (RAG), and modern agentic AI systems.

What began as an effort to learn Python for teaching and engineering applications gradually evolved into a deeper exploration of how intelligent systems reason, retrieve information, collaborate, critique their own outputs, and learn from past decisions. Every project in this portfolio represents not only a completed system, but also the questions, experiments, debugging sessions, architectural redesigns, and lessons learned along the way.

My long-term goal is to bridge Artificial Intelligence and Engineering, applying machine learning, NLP, generative AI, and agentic systems to manufacturing, engineering education, and real-world decision support. This portfolio showcases that journey through practical AI systems, continuous learning and experimentation.

SNS College of Technology · AICTE QIP Programme · IIIT Allahabad . NPTEL Course Instructors · Kaggle · UCI · Hugging Face · CrewAI · LangGraph · LangChain · PyTorch · TensorFlow · Open-Source AI Communities