Things I've Built
& Am Building.

A production-grade 7-stage RAG pipeline for research document intelligence — grounded answers with inline source and page citations across dense business document corpora.

A LangChain agentic application that generates structured learning plans using a ReAct agent (create_agent) with custom tools for course lookup and completion-time calculation. Deterministic logic, validation and calculations, runs in Python; the LLM handles tool selection and plan generation. Includes an LCEL-based Explain mode and a Streamlit interface.

A multi-node LangGraph workflow that takes a stock ticker and an optional question, combines yfinance price indicators (SMA20, SMA50, volatility, trend) with a RAG pass over roughly 14 days of NewsAPI headlines, and returns a simulated BUY, SELL or HOLD with a justification. A conditional edge routes invalid tickers and API errors to a fallback node. Informational only: no trades are executed and it is not financial advice.

Eliminates the 30–45 min of manual research before an HR screening call. Upload your resume, paste the job description, and enter the company name — a Claude API call (routed through a Cloudflare Worker proxy, so no API key needed) pulls the last 6 months of company news, funding rounds, and product launches, maps your skills to what the company needs, generates personalised talking points, and produces 3 smart questions to ask HR — structured company and role intelligence in under 60 seconds. Listed on Gumroad at $9.

An automated morning briefing built on n8n — RSS feeds across AI/ML, tech, job market, and global news are pulled on a schedule, synthesised into a readable digest by an LLM, and delivered straight to your inbox. Zero manual effort after setup.

Built from scratch — no frameworks, no templates, no shortcuts. A fully responsive dark-theme portfolio showcasing strategy, analytics, and AI work. Custom CSS animations, smooth scroll navigation, and a design where every section and pixel was deliberate. Deployed on GitHub Pages.

A RAG-based conversational app built in two stages. First, implemented the full retrieval pipeline manually: PDF extraction, chunking, sentence-transformer embeddings, ChromaDB vector storage, and grounded prompt construction. Then rebuilt the same pipeline using LangChain's RetrievalQA chain to directly compare manual wiring against framework abstraction and understand what the framework is actually doing under the hood. Answers are generated via Groq's LLaMA 3.3 70B and grounded strictly in the source document. Tested on a 139-page financial report with zero hallucination.

An AI-powered resume screening tool built with a deliberate architectural decision at its core. Evaluated a RAG/LangChain pipeline first, then chose direct context injection after confirming both documents fit comfortably within the LLM's context window, avoiding unnecessary complexity. Upload a resume and a job description and get an instant match score out of 100, a breakdown of matching and missing skills, and AI-generated recommendations to close the gap. Structured prompt engineering returns consistent JSON output every time. Supports PDF, Word, and plain-text inputs via pdfplumber and python-docx. Deployed as an interactive Gradio app on Hugging Face Spaces with a cost-controlled demo mode.

A hands-on fine-tuning lab applying SFT, LoRA, and QLoRA to a domain-specific research task: measuring accuracy, inference latency, GPU memory, and model size at each adaptation stage.