Making data speak clearly.

RAG chatbot running 100% locally: BGE-M3 embeddings (dense + lexical) stored in a Qdrant vector database feed a Qwen2.5-1.5B SLM served by Ollama — the lowest-latency pick from four stress-tested local models — to answer UCD academic and admin policy questions, with mandatory citation (URL + page number) and answers grounded in retrieved context (the model abstains when the docs do not support an answer). Built a scraping and parsing pipeline with BeautifulSoup, markdownify, and PyMuPDF (pymupdf4llm) to convert 200+ policy docs to structured Markdown, preserving hierarchical structure for downstream vector indexing. Automated dedup and validation checks cut malformed records by ~30% and pushed retrieval precision above 80%.

Tracks visa decisions published daily by the Embassy of Ireland, New Delhi, turning a flat daily spreadsheet (no history, no search, no trends) into a searchable, browsable dashboard. A GitHub Actions scraper runs 4× daily to fetch the embassy’s ‘.ods’ file, handle weekends and public holidays gracefully, and upsert records into a Google Sheet (the database). The Apps Script Web App exposes a clean HTTP API read by a four-tab dashboard: live snapshot, daily summary, full history with date-picker, and a suggestions form.

Built an end-to-end telemetry analytics pipeline using Python, FastF1, pandas, and NumPy to collect, clean, transform, and analyse race telemetry across the 2026 Formula 1 season, publishing weekly data-driven performance reports. Developed 9+ visualisation types including lap-time evolution, tyre strategy analysis, minisector dominance, cumulative lap delta, and sector performance, translating complex datasets into audience-friendly visual stories through 10+ published articles on Medium.

Cleaned and analysed seven real-world datasets — bank churn, US airline on-time performance, Bangalore BMTC transit, pizza sales, sleep health, Lufthansa, and an IBM Coursera fleet inventory — in Excel and pandas, fixing missing values, outliers, and type mismatches before touching a chart. Tableau charts surface the correlations worth reporting; every workbook and notebook lives in the repo.

Two dashboards published on Tableau Public: a customer analysis dashboard and a CO2 emissions view. Filters, parameters, and calculated fields let viewers explore the data themselves rather than just reading a chart — the customer analysis dashboard has collected 24 loves from the Tableau community.

Developed a Convolutional Neural Network to classify eye images as healthy or infected with conjunctivitis. Implemented Conv2D, MaxPooling, ReLU, Dropout, and Dense layers; trained on a labeled dataset of healthy and infected eyes. Integrated with a Flask app for real-time camera input, allowing users to self-screen and take timely action.

Built a privacy-first personal expense tracker with no framework and no build step. Just vanilla ES modules. Type a plain-language entry like “lidl groceries 4.47 revolut” and it auto-parses amount, category, and payment method. Data stays in localStorage with optional private GitHub sync, secured by a PIN or Face ID / fingerprint lock.