Building Across the AI Stack: RAG, LangChain, LangGraph Agents, LLM Fine-Tuning and Distributed Inference
Building Across the AI Stack: RAG, LangGraph Agents, LLM Fine-Tuning and Distributed Inference
AI engineering involves more than connecting an LLM to an API. Building reliable AI applications requires an understanding of retrieval, model behavior, training techniques, orchestration, and inference infrastructure.
I've organized nine public projects to explore these areas through source code, implementation examples, and technical documentation.
Portfolio: https://su394611-del.github.io/
1. Retrieval-Augmented Generation and Information Retrieval
Retrieval systems introduce architectural decisions that affect relevance, security, and response quality.
Some important design considerations include:
- Selecting embedding models and vector databases.
- Maintaining tenant-specific retrieval boundaries.
- Designing retrieval APIs with FastAPI.
- Implementing caching without compromising data isolation.
- Evaluating retrieval quality independently of language generation.
My RAG Service explores FastAPI, Qdrant, tenant-scoped retrieval, and semantic caching.
A particularly important distinction is between retrieving relevant information and generating grounded answers. Evaluate retrieval quality before attributing improvements to the language model itself.
2. Agentic AI with LangGraph
AI agents need more than prompting. They require controlled execution, explicit state management, tool integration, and predictable routing.
My LangGraph project explores concepts such as StateGraph, MessagesState, state reducers, Pydantic validation, and reusable subgraphs.
For agentic systems, I consider three questions especially important:
- How should state move between agents?
- How can tool inputs and outputs be validated?
- How should a workflow handle invalid outputs and execution failures?
Graph-based orchestration can make these execution paths easier to reason about and inspect.
3. LLM Training and Fine-Tuning
Training and adapting language models introduce trade-offs involving GPU memory, optimization methods, dataset quality, and evaluation.
The projects explore:
- Transformer attention mechanisms and encoder architectures.
- Supervised fine-tuning and LoRA.
- Preference optimization approaches such as DPO and ORPO.
- DistilBERT fine-tuning configurations.
- Distributed training using PyTorch DDP.
The LLM Fine-Tuning project provides implementation examples of model adaptation and preference optimization techniques.
An essential engineering principle is that successful code execution does not automatically establish model quality. Training results need to be evaluated using appropriate datasets, metrics, and reproducible configurations.
4. Distributed Training and Parallel Inference
Training and inference present different scaling challenges.
DistributedDataParallel typically synchronizes model gradients during training, while multi-worker inference often uses independent model replicas to process separate workloads.
The portfolio explores both approaches:
Important measurement areas include GPU utilization, model-loading overhead, throughput, latency, and memory consumption.
Implementation examples and performance-validated systems should be distinguished clearly, especially when physical multi-GPU execution has not been demonstrated.
5. Other Projects in the Collection
The portfolio also includes a Transformer encoder implementation, DistilBERT training comparisons, Transformer evaluation workflows, and an MCP/LangGraph market simulation.
Together, these projects examine different parts of the AI engineering lifecycle rather than treating LLM integration as a single problem.
6. What I Want to Explore Next
The areas I find especially valuable for further technical investigation include reproducible model evaluation, retrieval benchmarking, agent reliability, GPU performance profiling, and systematic comparisons between training configurations.
My goal is to make the implementations easier to inspect, reproduce, test, and improve.
Explore the Projects
All nine project websites, GitHub repositories, and supporting documentation are linked from the central portfolio:
https://su394611-del.github.io/
I welcome technical feedback, architecture discussions, bug reports, and collaboration from other developers working with modern AI systems.