AI Stack

AI Engineering Stack

The tools, frameworks, and concepts I build with — not just use.

Anthropic Certified
🧠Research LLM
Claude (Anthropic)
Primary LLM for production builds. Used via the Claude API for the Pre-Call Intel Briefing Tool and the Daily Intelligence Digest, with hands-on experience in prompt grounding, structured JSON output, and cost-controlled demo deployments.
Certified ✓
💬LLM
ChatGPT & Gemini
Multi-model workflows for research synthesis, content generation, and cross-model comparison during development and testing.
Daily use
🔍AI Search
Perplexity & NotebookLM
Real-time research and source-grounded analysis. Used for market intelligence gathering and deep document research during the 1Lattice consulting engagement.
Research stack
🧬Frameworks
LangChain & LangGraph
Core framework for building LLM applications and agentic pipelines. Used in the RAG PDF Chat app and the Resume vs JD Gap Analyser — implemented manual pipelines first, then rebuilt with LangChain to understand the abstraction layer.
Building with
🗄️Vector & Retrieval
ChromaDB & Pinecone
Vector databases for embedding storage and semantic retrieval. Used ChromaDB locally in RAG pipelines with sentence-transformer embeddings, and exploring Pinecone for production-scale deployments.
RAG stack
⚡LLM Inference
Groq
Fast inference API for running Llama and Mistral models. Used in the RAG PDF Chat app for low-latency responses during development and testing.
Fast inference
🚀Deployment
Streamlit & Gradio
Rapid UI and deployment layer for Python-based AI apps. Deployed the RAG PDF Chat on Streamlit Community Cloud and the JD Gap Analyser on Hugging Face Spaces via Gradio.
Shipped and live
⚙️Automation
n8n
Workflow automation tool used to build the Daily Intelligence Digest: RSS aggregation, Claude Haiku summarisation, and scheduled Gmail delivery across 50+ sources daily.
In production
🤝Agent Architecture
MCP (Model Context Protocol)
Hands-on with Model Context Protocol for building tool-augmented agent architectures. Practical exposure to connecting LLMs with external tools and structured data sources.
Hands-on
🧲Embeddings
Sentence Transformers & Hugging Face
Used sentence-transformer models for generating embeddings in RAG pipelines. Deployed the JD Gap Analyser on Hugging Face Spaces with a cost-controlled demo mode.
In pipeline
📄PDF Processing
PyMuPDF
Used for coordinate-based PDF text extraction in the Deep Research RAG pipeline. Handles dense financial documents including annual reports and AGM transcripts reliably across 444 pages.
In production
⚖️Evaluation
RAGAS
Local evaluation framework for RAG pipelines. Used to evaluate Deep Research RAG on 13 ground-truth QA pairs across four metrics: Context Precision 0.991, Context Recall 1.000, Faithfulness, and Answer Relevancy. Runs entirely locally with no OpenAI API calls.
Evaluated
🐳Containerisation
Docker
Used to containerise the Deep Research RAG system with ChromaDB baked into the image and Hugging Face model weights cached in a named Docker volume so they download once and persist across container restarts.
Shipped
🚂Deployment
Railway
Cloud deployment platform used to host the Dockerised Deep Research RAG FastAPI application. Chosen after Hugging Face Spaces moved to paid Docker tier and Render hit out-of-memory errors on the free tier.
Live
🔧Agentic AI
ReAct Agents & Tool Calling
Built a ReAct agent (LangChain create_agent) with custom tools for deterministic calculations and data lookup, in the Learning Sprint Planner Agent. Tested tool-grounding failure modes, where an LLM attempted to answer despite a tool returning no data, and enforced authoritative validation outside the LLM as a result.
Built & tested
Conceptual Foundations
Transformer Architecture Fine-tuning Embeddings and Tokenization RAG (Retrieval-Augmented Generation) LLM Evaluation and Guardrails Prompt Engineering Deep Learning Fundamentals Bias-Variance Tradeoff Supervised and Unsupervised Learning Vector Databases Agent Architectures Retrieval Engineering Hybrid Search BM25 Sparse Retrieval Tool Calling ReAct Pattern Agent Grounding